fix: #1876 LiteLLM extra_body forwarding

[yu2001-s]

Summary

• pass ModelSettings.extra_body via LiteLLM's extra_body kwarg instead of flattening its keys into top-level litellm.acompletion() kwargs
• keep reasoning_effort promotion to the top-level LiteLLM parameter while removing the duplicated key from the forwarded extra_body copy
• add regression tests for nested extra_body, nested metadata, and reasoning-effort precedence behavior

Why

LiteLLM already supports extra_body and merges it into the upstream OpenAI-compatible request body. Flattening ModelSettings.extra_body in the SDK changes the request shape and can break proxy-specific fields like nested metadata, session tags, and cached-content hints.

Validation

• uv run pytest -q tests/models/test_litellm_extra_body.py
• uv run ruff check src/agents/extensions/models/litellm_model.py tests/models/test_litellm_extra_body.py
• uv run mypy src/agents/extensions/models/litellm_model.py tests/models/test_litellm_extra_body.py
• make format
• make lint
• make typecheck currently fails on an unrelated mypy error in tests/test_run_step_execution.py:1231 (Module has no attribute "eager_task_factory") outside this diff

Fixes #1876

[seratch]

@yu2001-s Thanks for sending this. I've confirmed this works way better than the current implementation. Once all the things on your side are ready, mark this PR as ready for review.