Replace socketpair write bench with deterministic partial-send microbench

[Kludex]

Summary

• The previous test_bench_sync_stream_write_large used a real socket.socketpair() and a draining reader thread. In practice the kernel accepted most sends in one go, so the buffer-slicing loop inside SyncStream.write ran only a handful of times. That meant optimisations targeting that loop (e.g. perf: use memoryview in write() to avoid copies #954) did not produce a detectable delta on CodSpeed.
• This PR replaces it with test_bench_sync_stream_write_microcopy, which drives SyncStream.write through a _PartialSendSocket that caps send() at 4KB per call. The Python-level slicing loop now runs ~1024 times against a 4MB payload, so any change to that loop is the dominant cost of the bench.
• Locally verified that flipping buffer = buffer[n:] to view = view[n:] (the perf: use memoryview in write() to avoid copies #954 change) moves this bench from ~13.8ms to <1µs - far above any walltime noise floor.

Test plan

• CodSpeed walltime CI runs cleanly on this PR
• Re-run perf: use memoryview in write() to avoid copies #954 against main with this bench in place and confirm the memoryview change shows up

AI Disclaimer

This PR was developed with the assistance of either Claude or Codex. I've reviewed and verified the changes.

[codspeed-hq]

Merging this PR will not alter performance

✅ 6 untouched benchmarks
🆕 1 new benchmark
⏩ 1 skipped benchmark¹

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
🆕	WallTime	test_bench_sync_stream_write_microcopy	N/A	511.8 ms	N/A

---

_{Comparing sensitize-sync-write-bench (b2713e5) with main (8d5f182)}

Footnotes

1. 1 benchmark was skipped, so the baseline result was used instead. If it was deleted from the codebase, click here and archive it to remove it from the performance reports. ↩