[codex] Honor max-connections above HTTP client cap

[ywang96]

Summary

• replace the raw httpx.AsyncClient in KimiAPI with OpenAIAsyncHttpxClient
• preserve the long streaming timeout while removing the hidden active-connection cap from the HTTP pool
• let inspect_ai's max_connections semaphore remain the effective concurrency control

Root Cause

--max-connections was passed through correctly, but the provider created a raw httpx.AsyncClient with httpx defaults. That client defaults to max_connections=100, so requests above 100 were throttled by the HTTP pool before they could honor the configured inspect-ai concurrency.

Validation

• uv run python -m py_compile kimi_model.py
• instantiated _default_http_client() and verified the transport pool no longer caps active connections at 100 while keeping max_keepalive_connections=100

[ywang96]

Without this change, evaluating with --max-connnections bigger than 100 will slightly fall back to 100 becasue of httpx default max connections.