Tune gunicorn lifecycle for streaming requests

[wilsonccccc]

Summary

• Move Gunicorn runtime settings into gunicorn_conf.py so lifecycle parameters are explicit and env-overridable.
• Increase request and graceful shutdown timeouts to 300s for long provider streams.
• Disable max_requests worker recycling by default to avoid interrupting streaming /v1/chat/completions requests.

Context

Railway HTTP logs showed many /v1/chat/completions 502s with upstream connection closed unexpectedly. The old deployment also showed repeated Maximum request limit ... Terminating process events, which can recycle workers while streaming requests are active.

Validation

• python3 -m py_compile gunicorn_conf.py
• Loaded gunicorn_conf.py locally and confirmed defaults: timeout=300, graceful_timeout=300, keepalive=75, max_requests=0, max_requests_jitter=0.
• git diff --check HEAD~1..HEAD