fix(llm): [cherry-pick 1.3.0] reject unsupported RL fields for responses

[jthomson04]

Summary

Cherry-pick #11253 into release/1.3.0 so Responses API requests consistently reject unsupported RL output fields before engine dispatch.

• Clean cherry-pick of merged commit 769c6fc51822089d8a155217ee93acafc51083f8; no conflicts.
• Rejects completion_token_ids and prompt_logprobs when requested through nvext.extra_fields on /v1/responses.
• Returns a field-specific 501 response for both streaming and non-streaming requests.
• Leaves Chat Completions, Completions, and supported Responses nvext fields unchanged.

Root cause

Responses requests are converted to the Chat Completions protocol internally. The non-streaming conversion copied the resulting chat response nvext verbatim, while the Responses streaming converter did not expose those fields. This unintentionally made the RL fields available only on the non-streaming Responses path.

Where should the reviewer start?

Review validate_response_unsupported_fields and its tests in lib/llm/src/http/service/openai.rs.

Related Issues

• Backport of fix(llm): reject unsupported RL fields for responses #11253
• Relates to DYN-3392

Validation

• cargo fmt --all -- --check
• cargo test -p dynamo-llm test_validate_unsupported_fields --lib (8 passed)
• cargo clippy -p dynamo-llm --lib -- -D warnings
• git diff --check refs/remotes/origin/release/1.3.0...HEAD

[devin-ai-integration]

✅ Devin Review: No Issues Found

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

[dagil-nvidia]

Release backport quality gate: exact source patch equivalence, targeted Responses API regression tests, and current CI verified at head decb8f2.