Fix OpenAI API request option handling

[patrick-andstar]

Summary

• Forward the existing max_length request option to the underlying chat generation call.
• Accept stop as either a string or a list of strings, and normalize a scalar before applying the existing leading-newline workaround.

Motivation

max_length was accepted by the request model but silently ignored. In addition, clients using the standard scalar form of stop received a validation error even though this script exposes an OpenAI-compatible chat-completions endpoint.

Compatibility

Existing list-valued stop input keeps its current behavior. No model loading, streaming protocol, or authentication behavior changes.

Testing

• Ran a focused stubbed endpoint regression check covering scalar stop, existing list stop, and forwarding max_length to model.chat.
• Ran an AST syntax check for openai_api.py.
• Ran git diff --check.
• Did not run the Docker/model integration test because it requires Docker and an external checkpoint download.

Duplicate Work Check

Searched open Issues and PRs for openai_api, max_length, and stop words; no matching open change was found. Historical PR #1038 added streaming stop-word support but does not cover these request options. Open PR #2255 changes the unrelated plugin-evaluation parser.