epic: close OpenAI API conformance gaps in Guardrails server

Goal

The NeMo Guardrails server's OpenAI-compatible API should match the
CreateChatCompletionRequest / CreateChatCompletionResponse / ListModels
contracts closely enough that standard OpenAI SDK clients work without
surprises. This is an important step for this server to become a transparent inference proxu

Why

The conformance CI (api-conformance-impl.yml) reports around 100 gaps between
app.openapi() and OpenAI v2.3.0. Some are harmless schema metadata diffs, but
others cause real problems like missing request properties, route metadata differences etc

Scope

Potential sub-issues might include:

1. Missing request properties (~23 gaps)
2. Request schema typing — nullable, defaults, anyOf/oneOf (~24 gaps)
3. Response usage required fields (~7 gaps)
4. Response types, enums, nullables, streaming media type (~24 gaps)
5. Route metadata — operation IDs, tags (4 gaps)
6. GET /models response schema (~9 gaps)
7. Conformance script noise reduction (~10 gaps)

Local reproduction (requires oasdiff):

# Install oasdiff e.g.
brew install oasdiff

# Run the conformance check (fetches the OpenAI spec and diffs via oasdiff)
uv run --locked python scripts/openai_coverage.py --fetch --fastapi

Completion criteria

• scripts/openai_coverage.py --fastapi reports 0 conformance gaps (excluding
  the intentional-extensions allowlist).
• Intentional Guardrails extensions (guardrails request/response properties,
  422 status) are reported separately, not as gaps.
• Unimplemented endpoints (Stored Completions, single-model CRUD) are excluded
  from the gap count.
• All existing tests pass; new tests cover the added request properties and
  response schema changes.

Start with scripts/openai_coverage.py and run uv run --locked python scripts/openai_coverage.py --fetch --fastapi to review the reported gaps against the OpenAI v2.3.0 contracts. Work through the listed request, response, route, and model-schema differences, then verify that the command reports no unallowlisted gaps and that all existing and new tests pass.