fix(models): allow extra_args when chat completions kwarg is omitted - #3200

Summary

Mirror of #3185 (Responses API fix) for the Chat Completions path.

When a user supplies a first-class request parameter (e.g. temperature, top_p, max_tokens, metadata, ...) through ModelSettings.extra_args without also setting the corresponding ModelSettings field, the chat completions path falsely raises:

TypeError: chat.completions.create() got multiple values for keyword argument 'temperature'

This happens because the field in create_kwargs resolves to OpenAI's omit sentinel (not absent), and the duplicate-key check uses a plain set intersection. The Responses API path had the same bug, fixed in #3185 by skipping keys whose first-class value is Omit/NotGiven. The same pattern still exists in openai_chatcompletions.py::_fetch_response — this PR applies the same fix here.

Behavior change

• Before: ModelSettings(extra_args={"temperature": 0.25}) raised TypeError unless temperature was also None-by-default normalization.
• After: the extra_args value passes through.
• Genuine collisions (e.g. ModelSettings(temperature=0.5, extra_args={"temperature": 0.25})) still raise TypeError as before.

Test plan

• uv run pytest tests/models/test_openai_chatcompletions.py tests/models/test_openai_chatcompletions_stream.py tests/models/test_openai_chatcompletions_converter.py tests/models/test_kwargs_functionality.py (59 passed)
• ruff check + ruff format --check clean
• New tests:
  • test_fetch_response_allows_extra_arg_when_explicit_arg_is_omitted — fails on main, passes after fix
  • test_fetch_response_rejects_duplicate_extra_args_with_explicit_value — guards regression on the genuine-collision case

Issue number

N/A (parallel of #3185 for chat completions)

Checks

• I've added new tests
• I've added/updated the relevant documentation (none needed)
• I've run make lint / make format equivalent (ruff check / ruff format --check)
• I've made sure tests pass