Add logprobs support to chat completions

[raullenchai]

The OpenAI API spec supports logprobs: true in chat completion requests, returning token log-probabilities in the response.

Currently Rapid-MLX accepts the parameter silently but doesn't return logprobs data. The response choices[].logprobs field is absent.

What needs to happen:

• When logprobs: true, capture per-token log-probabilities from mlx-lm's generate_step()
• Return them in the choices[].logprobs field per the OpenAI spec
• Add a test case

Relevant files: vllm_mlx/server.py (chat completion endpoint)

[raullenchai]

Already implemented. Both streaming and non-streaming chat completion paths support logprobs: true + top_logprobs: N. Uses _extract_token_logprob() to collect per-token log-probabilities from mlx-lm's generate_step() and returns them in choices[].logprobs.content[] per OpenAI spec. See routes/chat.py lines 429-456 (non-streaming) and 620-632 (streaming).

[raullenchai]

Reopening — investigated further. The route-level logprobs code exists (extraction + formatting) but BatchedEngine.stream_chat() does not return logprobs data (output.logprobs is always None, output.tokens is always []). The engine layer needs to propagate logprobs from mlx-lm's generate_step() through the scheduler to GenerationOutput. This is an engine-level change, not just a route-level one.