perf: reduce async request overhead

[WH-2099]

Closes #1182.

Discussion: #1181.

Summary

This drops three bits of avoidable work from common request paths:

• Skip the cancellation shield when the connection pool has nothing to close.
• Reuse client headers when a request does not add any.
• Pass unencoded async response chunks through without no-op chunking work.

The changes are split into three commits so each one can be reviewed or reverted on its own.

Performance

Scenario	Requests/s	Client CPU/request
c1 / 1 KiB	4,710 -> 4,994 (+6.0%)	210.0 -> 198.1 us (-5.7%)
c32 / 1 KiB	4,568 -> 4,813 (+5.4%)	217.2 -> 206.2 us (-5.1%)
c32 / 64 KiB	3,941 -> 4,114 (+4.4%)	251.7 -> 241.1 us (-4.2%)

These are medians from five interleaved 2.5-second runs on CPython 3.14.7 with stdlib asyncio and pinned server/client CPUs.

The benchmark used plain HTTP/1.1 on loopback, so the numbers should not be read as TLS, HTTP/2, or WAN results.

Validation

• scripts/test
• 1995 passed, 1 skipped
• 100% coverage
• Ruff, mypy, lockfile, and unasync checks pass

Checklist

• I understand that this PR may be closed in case there was no previous discussion.
• I've added regression tests for the behavior-sensitive changes and kept the commits atomic.
• No documentation update is needed for these internal changes.

[codspeed-hq]

Merging this PR will not alter performance

✅ 17 untouched benchmarks
⏩ 7 skipped benchmarks¹

---

_{Comparing WH-2099:perf/reduce-async-request-overhead (42a9350) with main (bd0b9bc)}

Footnotes

1. 7 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[cubic-dev-ai]

No issues found across 6 files

_{Re-trigger cubic}

[WH-2099]

@Kludex Could you take a look at this? (⌐■_■)