feat(openai): expose max_output_tokens on Responses API LLM

[piyush-gambhir]

Summary

Adds max_output_tokens as a first-class constructor argument on the OpenAI, Azure, and xAI Responses API LLM wrappers — mirroring how service_tier is handled today (#5346).

The underlying /v1/responses endpoint accepts max_output_tokens natively (see openai.types.responses.response_create_params), but the plugins previously required callers to reach for extra_kwargs on every chat() call to set it. This PR makes it a plain constructor arg so a bound can be set once at LLM construction time.

Changes

• livekit-plugins-openai/.../responses/llm.py: add max_output_tokens: NotGivenOr[int] to _LLMOptions and the LLM.__init__ signature; forward it into extra so it reaches both the HTTP (client.responses.create) and WebSocket (response.create) paths.
• livekit-plugins-azure/.../responses/llm.py: thread the new arg through to openai.responses.LLM.__init__.
• livekit-plugins-xai/.../responses/llm.py: thread the new arg through to openai.responses.LLM.__init__.

Usage

from livekit.plugins import openai

llm = openai.responses.LLM(
    model="gpt-4.1",
    max_output_tokens=1024,
)

Motivation

All sibling chat-completions LLMs (openai.LLM, anthropic.LLM, google.LLM, aws.LLM, mistralai.LLM, groq.LLM) already expose a first-class max_completion_tokens / max_tokens / max_output_tokens argument. The Responses variants were the only remaining path that forced callers into extra_kwargs. This brings them in line.

Test plan

• ast.parse passes on all three modified files
• make check (format + lint + type-check)
• Smoke test: instantiate openai.responses.LLM(model="gpt-4.1", max_output_tokens=64) and verify the response honors the cap in both use_websocket=True and use_websocket=False modes

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[piyush-gambhir]

Smoke test (live OpenAI API, gpt-4.1):

• HTTP path ✅: LLM(model='gpt-4.1', use_websocket=False, max_output_tokens=16) — 500-word-story prompt returned truncated output (~12–16 tokens, e.g. "Once upon a time, in the misty blue mountains of Elderglow,"). Confirms max_output_tokens=16 is forwarded to client.responses.create(...) and honored by the API.
• WebSocket path ✅: stream._extra_kwargs['max_output_tokens'] = 16 before send — the param is spread into the response.create payload dict via the same **self._extra_kwargs path that carries every other option (temperature, service_tier, etc.), so wiring is symmetric with HTTP.

make check passes locally (only pre-existing warnings, none in changed files).