fix(evals): avoid overriding judge reasoning

[inickt]

Summary

• Stop the generic eval judge from injecting the Chat Completions-specific reasoning_effort="none" option.
• Preserve reasoning options configured on the supplied LLM.
• Add regression coverage for the options forwarded by the judge.

Why

The judge accepts any LLM, but reasoning_effort is specific to Chat Completions. Passing it to an OpenAI Responses client fails because Responses uses the reasoning object instead.

Explicitly configured reasoning options continue to apply. LLMs without an explicit configuration will now use their provider or model default rather than having the judge force reasoning off.

Testing

• Added focused unit coverage.
• Ran the full formatting, linting, and type-check suite.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[tinalenguyen]

thank you for the PR, this change makes sense. i think maybe we should set the default reasoning for inference GPT 5 models to none before applying this change so the judges don't suddenly start with medium effort. i can make that PR