vtprotobuf: Reuse backing array with referenced structs instead of pooling individual objects

[carsonip]

Alternative to #37 , but using vtprotobuf fork that reuses backing array including referenced structs, which has an upstream PR. With this, we will no longer be pooling Keyed* individually. Either with this or #37, there will be significantly fewer garbage.

Closes #20

benchstat:

goos: linux
goarch: amd64
pkg: github.com/elastic/apm-aggregation/aggregators
cpu: 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz
                                       │ bench-out/vt-unmarshal-pooling-including-structs-2-before │ bench-out/vt-unmarshal-pooling-including-structs-2-after │
                                       │                          sec/op                           │              sec/op                vs base               │
AggregateCombinedMetrics-16                                                            2.001µ ± 1%                         1.926µ ± 1%   -3.75% (p=0.002 n=6)
AggregateBatchSerial-16                                                                5.227µ ± 1%                         4.489µ ± 5%  -14.11% (p=0.002 n=6)
AggregateBatchParallel-16                                                              5.524µ ± 9%                         5.008µ ± 3%   -9.33% (p=0.002 n=6)
CombinedMetricsEncoding-16                                                            1410.0n ± 3%                         878.5n ± 2%  -37.69% (p=0.002 n=6)
CombinedMetricsToBatch-16                                                              61.90µ ± 1%                         61.41µ ± 2%        ~ (p=0.132 n=6)
EventToCombinedMetrics-16                                                              420.7n ± 2%                         418.4n ± 3%        ~ (p=0.240 n=6)
NDJSONSerial/go-2.0.0.ndjson-16                                                        19.11m ± 3%                         16.84m ± 4%  -11.88% (p=0.002 n=6)
NDJSONSerial/nodejs-3.29.0.ndjson-16                                                  10.980m ± 4%                         9.488m ± 3%  -13.59% (p=0.002 n=6)
NDJSONSerial/python-6.7.2.ndjson-16                                                    27.13m ± 4%                         23.21m ± 3%  -14.45% (p=0.002 n=6)
NDJSONSerial/ruby-4.5.0.ndjson-16                                                      16.14m ± 1%                         13.95m ± 4%  -13.57% (p=0.002 n=6)
NDJSONParallel/go-2.0.0.ndjson-16                                                      19.17m ± 4%                         16.71m ± 7%  -12.85% (p=0.002 n=6)
NDJSONParallel/nodejs-3.29.0.ndjson-16                                                11.005m ± 3%                         9.518m ± 6%  -13.51% (p=0.002 n=6)
NDJSONParallel/python-6.7.2.ndjson-16                                                  27.26m ± 2%                         23.03m ± 3%  -15.54% (p=0.002 n=6)
NDJSONParallel/ruby-4.5.0.ndjson-16                                                    16.25m ± 3%                         13.98m ± 3%  -13.97% (p=0.002 n=6)
geomean                                                                                458.6µ                              399.0µ       -13.01%

                                       │ bench-out/vt-unmarshal-pooling-including-structs-2-before │ bench-out/vt-unmarshal-pooling-including-structs-2-after │
                                       │                           B/op                            │             B/op              vs base                    │
AggregateCombinedMetrics-16                                                        1.416Ki ±  0%                     1.104Ki ± 1%   -22.01% (p=0.002 n=6)
AggregateBatchSerial-16                                                            2.459Ki ±  2%                     1.373Ki ± 3%   -44.17% (p=0.002 n=6)
AggregateBatchParallel-16                                                          2.500Ki ±  3%                     1.370Ki ± 3%   -45.21% (p=0.002 n=6)
CombinedMetricsEncoding-16                                                           855.0 ± 44%                         0.0 ± 0%  -100.00% (p=0.002 n=6)
CombinedMetricsToBatch-16                                                          3.986Ki ±  3%                     3.982Ki ± 0%         ~ (p=0.502 n=6)
EventToCombinedMetrics-16                                                            0.000 ±  0%                       0.000 ± 0%         ~ (p=1.000 n=6) ¹
NDJSONSerial/go-2.0.0.ndjson-16                                                   11.591Mi ±  0%                     9.249Mi ± 1%   -20.21% (p=0.002 n=6)
NDJSONSerial/nodejs-3.29.0.ndjson-16                                               7.007Mi ±  0%                     5.589Mi ± 1%   -20.23% (p=0.002 n=6)
NDJSONSerial/python-6.7.2.ndjson-16                                                19.59Mi ±  1%                     15.32Mi ± 1%   -21.78% (p=0.002 n=6)
NDJSONSerial/ruby-4.5.0.ndjson-16                                                 10.418Mi ±  0%                     8.261Mi ± 0%   -20.71% (p=0.002 n=6)
NDJSONParallel/go-2.0.0.ndjson-16                                                 11.597Mi ±  0%                     9.233Mi ± 1%   -20.38% (p=0.002 n=6)
NDJSONParallel/nodejs-3.29.0.ndjson-16                                             7.011Mi ±  0%                     5.563Mi ± 1%   -20.66% (p=0.002 n=6)
NDJSONParallel/python-6.7.2.ndjson-16                                              19.63Mi ±  1%                     15.32Mi ± 1%   -21.94% (p=0.002 n=6)
NDJSONParallel/ruby-4.5.0.ndjson-16                                               10.424Mi ±  0%                     8.256Mi ± 0%   -20.80% (p=0.002 n=6)
geomean                                                                                          ²                                 ?                      ² ³
¹ all samples are equal
² summaries must be >0 to compute geomean
³ ratios must be >0 to compute geomean

                                       │ bench-out/vt-unmarshal-pooling-including-structs-2-before │ bench-out/vt-unmarshal-pooling-including-structs-2-after │
                                       │                         allocs/op                         │            allocs/op             vs base                 │
AggregateCombinedMetrics-16                                                           21.00 ± 0%                          17.00 ± 0%  -19.05% (p=0.002 n=6)
AggregateBatchSerial-16                                                               26.00 ± 4%                          13.00 ± 8%  -50.00% (p=0.002 n=6)
AggregateBatchParallel-16                                                             26.50 ± 6%                          13.00 ± 8%  -50.94% (p=0.002 n=6)
CombinedMetricsEncoding-16                                                            0.000 ± 0%                          0.000 ± 0%        ~ (p=1.000 n=6) ¹
CombinedMetricsToBatch-16                                                             95.00 ± 0%                          95.00 ± 0%        ~ (p=1.000 n=6) ¹
EventToCombinedMetrics-16                                                             0.000 ± 0%                          0.000 ± 0%        ~ (p=1.000 n=6) ¹
NDJSONSerial/go-2.0.0.ndjson-16                                                      142.1k ± 0%                         114.1k ± 2%  -19.66% (p=0.002 n=6)
NDJSONSerial/nodejs-3.29.0.ndjson-16                                                 79.18k ± 1%                         63.79k ± 1%  -19.43% (p=0.002 n=6)
NDJSONSerial/python-6.7.2.ndjson-16                                                  215.4k ± 1%                         162.7k ± 1%  -24.49% (p=0.002 n=6)
NDJSONSerial/ruby-4.5.0.ndjson-16                                                   124.53k ± 0%                         98.87k ± 0%  -20.61% (p=0.002 n=6)
NDJSONParallel/go-2.0.0.ndjson-16                                                    142.0k ± 0%                         113.9k ± 1%  -19.81% (p=0.002 n=6)
NDJSONParallel/nodejs-3.29.0.ndjson-16                                               79.20k ± 0%                         63.17k ± 1%  -20.24% (p=0.002 n=6)
NDJSONParallel/python-6.7.2.ndjson-16                                                215.4k ± 2%                         163.2k ± 1%  -24.23% (p=0.002 n=6)
NDJSONParallel/ruby-4.5.0.ndjson-16                                                 124.54k ± 1%                         99.09k ± 0%  -20.43% (p=0.002 n=6)
geomean                                                                                          ²                                    -22.21%               ²
¹ all samples are equal
² summaries must be >0 to compute geomean

macro benchmark:

                           │ github/run82.txt │         github/run83-vt.txt         │
                           │    events/sec    │  events/sec   vs base               │
10000AggregationGroups-512       13.76k ± 35%   15.90k ± 32%        ~ (p=0.132 n=6)
1000Transactions-512             13.72k ± 33%   18.83k ± 28%        ~ (p=0.065 n=6)
AgentAll-512                     9.124k ± 19%   8.998k ± 50%        ~ (p=0.818 n=6)
AgentGo-512                      11.15k ± 22%   12.01k ± 53%        ~ (p=0.310 n=6)
AgentNodeJS-512                  5.777k ± 28%   6.611k ± 29%        ~ (p=0.093 n=6)
AgentPython-512                  6.825k ± 26%   7.422k ± 76%        ~ (p=0.180 n=6)
AgentRuby-512                    7.332k ± 44%   7.387k ± 67%        ~ (p=0.368 n=6)
geomean                          9.193k         10.23k        +11.28%

[axw]

Would like to see the reslicing/indexing to the end of the slices tidied up, but otherwise LGTM!