Publish - avoid async state machine when possible

[caleblloyd]

Brings back optimizations from prior to #346

• Logic to skip async state machine when possible
• Locks switched back from channel-based to semaphore-based, but it appears they are faster
• Selective send memory consolidation
• Bigger send buffer sizes

Performed some benchmarks on Linux:

Send 10 Million Messages Serially:

main:

Completed: 9999701, Awaited: 299

1,412,795 msgs/sec ~ 172.46 MB/sec

This PR:

Completed: 9999980, Awaited: 20

2,688,705 msgs/sec ~ 328.21 MB/sec

So around 1.9x more throughput.

Publish Serial Microbench

main:

Method	Iter	Mean	Error	StdDev	Allocated
PublishAsync	64	126.0 μs	246.1 μs	13.49 μs	360 B
PublishAsync	512	584.0 μs	2,328.1 μs	127.61 μs	333 B
PublishAsync	1024	839.2 μs	365.5 μs	20.03 μs	358 B

This PR:

Method	Iter	Mean	Error	StdDev	Allocated
PublishAsync	64	127.1 μs	258.8 μs	14.19 μs	321 B
PublishAsync	512	377.8 μs	677.4 μs	37.13 μs	336 B
PublishAsync	1024	750.7 μs	1,845.1 μs	101.14 μs	339 B

Looks pretty similar

Publish Parallel Microbench

main:

Method	Concurrency	Mean	Error	StdDev	Gen0	Gen1	Allocated
PublishParallelAsync	1	1.708 s	4.9311 s	0.2703 s	10000.0000	-	30.52 MB
PublishParallelAsync	2	1.486 s	1.6187 s	0.0887 s	17000.0000	2000.0000	105.66 MB
PublishParallelAsync	4	2.079 s	0.9243 s	0.0507 s	79000.0000	2000.0000	256.4 MB

This PR:

Method	Concurrency	Mean	Error	StdDev	Gen0	Allocated
PublishParallelAsync	1	1.112 s	1.2659 s	0.0694 s	10000.0000	30.52 MB
PublishParallelAsync	2	1.426 s	2.9107 s	0.1595 s	61000.0000	180.45 MB
PublishParallelAsync	4	2.096 s	0.4181 s	0.0229 s	216000.0000	640.77 MB

Faster at concurrency = 1, similar timing at 2 and 4. Allocations jump in this PR quite a bit at concurrency = 4, although they don't escape Gen0. Might be worth investigating why.

[caleblloyd]

Marking as draft as there are quite a few failing tests I need to look into before it's ready for review

[mtmk]

Semaphore here is giving us better performance clearly. Main gain is probably not having to manage timeouts with cancellation tokens.

[caleblloyd]

Could also be seeing gains from skipping the async state machinery

[caleblloyd]

Ready for review

[mtmk]

LGTM