perf: Rewrite _assign_requests_to_connections as a single-pass loop

[mbeijen]

The previous implementation re-scanned self._connections per queued request to rebuild available_connections and idle_connections lists, giving O(N*M) behaviour when many requests are queued against a populated pool. The keepalive-eviction step in pass 1 was also quadratic, since sum(... is_idle() ...) was recomputed for every connection.

This rewrite:

• Walks self._connections once to drop closed connections and schedule expired ones for close.
• Counts idle connections once, then evicts surplus idle connections in a single pass to enforce max_keepalive_connections.
• Snapshots available_connections once before the queued-request loop, and consults it (rather than re-filtering self._connections) for each request.

Behaviour is preserved exactly:

• The same connection may still be assigned to multiple HTTP/1.1 requests in a single call. The pool's existing ConnectionNotAvailable retry loop in handle_async_request handles that case as before.
• HTTP/2 multiplexing onto a single connection is preserved for the same reason — connections are not removed from available_connections on assignment.

Inspired by encode/httpcore#1035 by @VictorPrins.

Citing some issues related to this, including in downstream libraries relying on httpcore:

• https://github.com/encode/httpx/issues/3215
• [Bug]: Cannot get past 50 RPS BerriAI/litellm#6592
• httpx client has very poor performance for concurrent requests compared to aiohttp openai/openai-python#1596
• Httpx not production ready under load encode/httpx#3100

[codspeed-hq]

Merging this PR will not alter performance

✅ 7 untouched benchmarks

---

_{Comparing mbeijen:perf/assign-requests-single-pass (80c9942) with main (b4c5940)}