conformance: add missing request properties to chat completion schema

Problem

GuardrailsChatCompletionRequest declares 14 of the 28+ properties in
OpenAI's CreateChatCompletionRequest. The remaining 23 are silently dropped,
breaking clients that send them.

High-impact missing properties include response_format, seed,
max_completion_tokens, n, stream_options, and reasoning_effort.

Full list: max_completion_tokens, response_format, seed, n,
top_logprobs, stream_options, user, reasoning_effort, service_tier,
store, metadata, audio, modalities, moderation, prediction,
safety_identifier, verbosity, web_search_options, prompt_cache_key,
prompt_cache_options, prompt_cache_retention, function_call, functions.

Proposed direction

Add all 23 fields as Optional[Any] = None to
GuardrailsChatCompletionRequest in server/schemas/openai.py and forward
them to generation_options.llm_params in chat_completion(), following the
existing pattern for max_tokens, temperature, etc.

Alternative: set model_config = ConfigDict(extra="allow") on the request
model to forward all unknown fields as llm_params automatically. Covers
future additions but loses explicit OpenAPI documentation.

Acceptance criteria

• All 23 properties are accepted by the request schema.
• Values are forwarded to the LLM provider via llm_params.
• Conformance gap count drops by 23.
• Existing tests pass; new tests confirm passthrough for representative fields.

Validation

• Run scripts/openai_coverage.py --fetch --fastapi and confirm the 23
  "removed the request property" gaps are resolved.
• Run make test TEST=tests/server/ to confirm no regressions.

Files

• nemoguardrails/server/schemas/openai.py
• nemoguardrails/server/api.py

Parent epic

This issue is a sub-issue of #2249:

epic: close OpenAI API conformance gaps in Guardrails server

Start in nemoguardrails/server/schemas/openai.py by comparing GuardrailsChatCompletionRequest with the listed missing properties, then inspect chat_completion() in nemoguardrails/server/api.py and the existing forwarding pattern. Add representative passthrough tests and run scripts/openai_coverage.py --fetch --fastapi followed by make test TEST=tests/server/; done means all 23 gaps are resolved without regressions.