fix: honor model request overrides at agent startup

[supplefrog]

Summary

• Load optional model.request_overrides from config.yaml during agent startup.
• Merge global request overrides into AIAgent.request_overrides without clobbering explicit per-turn overrides.
• Reuse the existing extra_body merge behavior so provider-native fields like OpenAI Responses text.verbosity can be configured without local source patches.

Refs #20203

Test Plan

• python -m pytest tests/agent/test_custom_provider_extra_body.py tests/run_agent/test_run_agent_codex_responses.py -q -o 'addopts=' — 83 passed, 1 warning

Notes

• Explicit per-turn request overrides still win over global model.request_overrides.
• This is intentionally a low-level escape hatch for provider-native request body fields that Hermes has not promoted to first-class config yet.

[supplefrog]

Closing because the first-class implementations for #20203 now cover the supported path directly. #29574 currently includes the provider capability boundary, nested text merge, validation, and gateway cache invalidation.

The generic model.request_overrides startup propagation may still be useful independently, but it should have its own focused issue rather than compete as an indirect solution to #20203.