feat: add context_length to /v1/models response

[lichengzhe]

Summary

Include context_length in the /v1/models endpoint response by reading max_position_embeddings from each model's config.json.

Motivation

Many OpenAI-compatible API gateways and proxy tools (e.g. one-api, new-api) read context_length from the /v1/models response to automatically determine a model's context window size. Without this field, users must manually configure context limits for each model, which is error-prone and tedious.

Related: #966 (KeyError on context_length)

Changes

• Added a get_context_length() helper that reads max_position_embeddings from a HuggingFace cached model's config.json
• Updated the HF cache model listing to include context_length when available
• Updated the local model path listing to also read and include context_length
• Handles models that nest the value under text_config (e.g. some vision-language models)
• Gracefully falls back (omits the field) if config.json is missing or unreadable

Example response

{
  "object": "list",
  "data": [
    {
      "id": "RepublicOfKorokke/Qwen3.5-35B-A3B-mlx-lm-mxfp4",
      "object": "model",
      "created": 1740000000,
      "context_length": 262144
    }
  ]
}

Testing

Tested locally with Qwen3.5-35B-A3B-mlx-lm-mxfp4 — confirmed context_length: 262144 appears in the response.