Optimize memo sync: lazy dirty flags instead of eager cloning

[jensens]

Summary

• Follow-up to #18 fix (819b21b) which introduced decode performance regression
• Replaces eager sync_memo_top() (clones full container after every SETITEMS/APPENDS) with lazy dirty-flag approach
• Memo entries marked dirty after mutations, resolved only when actually read via BINGET or when the owning stack slot is popped

Benchmark results (PGO build, 5000 iterations)

Decode median (us)

Category	Baseline	Eager (#18 fix)	Lazy (this PR)	vs Baseline
simple_flat_dict	1.0	1.5	1.3	+30%
nested_dict	1.6	3.2	2.5	+56%
large_flat_dict	18.0	31.5	25.1	+39%
special_types	3.8	6.8	5.6	+47%
wide_dict	244.5	585.8	348.6	+43%
deep_nesting	6.4	32.4	20.8	+225%

FileStorage (1,692 real records)

Metric	Baseline	Eager	Lazy
Decode median	23.6 us	43.7 us	34.5 us
Encode median	4.0 us	4.3 us	4.1 us

Recovers ~half the regression. Remaining overhead is inherent to value semantics — when a dirty memo'd container is popped from the stack, we must clone it into memo. Only reference semantics (Rc) could eliminate this, but that was previously tried and regressed worse overall.

Test plan

• cargo test — 200 tests pass (incl. 2 regression tests from #18)
• pytest tests/ — 182 Python tests pass
• PGO benchmark comparison against baseline

🤖 Generated with Claude Code