OpenAI backend token logprobs propagation into responses

[Zigfreidish]

Child of #1384.

Replace the current explicit unsupported logprobs receipt with real token-logprob propagation from the backend into the logprobs/top_logprobs response fields, when the runtime reports them. Runtime-dependent.

Done when

• When the backend reports token logprobs, they appear in the response with OpenAI-compatible shape.
• When it does not, the explicit unsupported receipt remains rather than synthesizing values.

[Akane-CN]

Watch note: backend token evidence should stay bytes-and-token faithful instead of being reconstructed after decoding.

• Design principle: batch generation APIs should optionally return generated token IDs and sampled-token logprobs alongside decoded text, with all three sequences kept length-aligned per request.
• Melix-specific goal: when a worker can report token IDs/logprobs, propagate them into the OpenAI response and into evaluation/RL evidence; when it cannot, keep the existing typed unsupported receipt rather than synthesizing values from text.
• Measurable target: for supported backends, 100% of emitted completion tokens have a matching token_id and logprob, lengths match the streamed/non-streamed text assembly fixtures, and unsupported backends still produce zero fabricated logprobs.
• Next executable slice: extend the worker response envelope and mock backend fixtures with optional completion_token_ids / completion_logprobs, then add one real-backend canary plus one conformance test that verifies OpenAI-compatible logprobs shape and fallback receipts.