Fix OpenAI completion args routing

[kayametehan]

Fixes #1493 and #1504.

Overview

--completion_args currently breaks in two ways for OpenAI completion functions:

1. OpenAIChatCompletionFn rejects extra kwargs such as temperature
2. OpenAICompletionFn accepts extra kwargs but silently drops them instead of forwarding them to the API request options

There is also a related registry-plumbing issue where registry can be passed into OpenAIChatCompletionFn when the class is referenced from the registry.

What changed

• Allow OpenAIChatCompletionFn to accept extra kwargs
• Promote extra constructor kwargs into extra_options for both OpenAI completion function classes
• Ignore registry plumbing kwargs instead of forwarding them to the API request
• Reuse numeric coercion for --completion_args, so values like temperature=0.5 and max_tokens=10 are parsed as numbers rather than strings
• Add focused regression tests for constructor behavior, registry routing, and CLI parsing

Validation

I verified this change before pushing with:

• python -m pytest evals/completion_fns/openai_test.py evals/cli/oaieval_test.py -q
• python -m pytest evals/registry_test.py evals/completion_fns/openai_test.py evals/cli/oaieval_test.py -q
• direct constructor / registry behavior checks for temperature and max_tokens
• git diff --check

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.