perf: use anyio fast_acquire for Lock and Semaphore

[mbeijen]

Reinstates encode/httpcore#953 which was reverted in encode/httpcore#1002 for process reasons, not technical ones. fast_acquire=True skips an unnecessary event loop yield on uncontended lock acquires, eliminating two round-trips per request through the connection pool.

Requires anyio >= 4.5.0 (bumped in pyproject.toml).

[codspeed-hq]

Merging this PR will not alter performance

✅ 7 untouched benchmarks

---

_{Comparing mbeijen:perf/fast-acquire-locks (ddc00d9) with main (79f788b)}