Include context_length in /v1/models response (#1183)

[seikixtc]

Closes #1183.

What

Adds a context_length field to every entry returned by the /v1/models endpoint (and the /v1/models/{repo_id} single-model variant), reporting the maximum context length the model declares in its config.json.

Example response after this change:

{
  "object": "list",
  "data": [
    {
      "id": "mlx-community/Qwen2.5-7B-Instruct-4bit",
      "object": "model",
      "created": 1745000000,
      "context_length": 32768
    }
  ]
}

Why

OpenAI-compatible clients increasingly need to know a model's usable context window before constructing a request — for chunking long documents, sizing prompt caches, choosing between models, and rendering token-budget UIs. Today the endpoint exposes only id, object, and created, so clients have to hard-code limits or guess.

How

• Adds _get_context_length(config_path) to read the first recognized max-context field from config.json
• Supports max_position_embeddings, n_positions, max_sequence_length, and seq_length
• Adds _find_repo_config_path(repo) so Hugging Face cache scans use the top-level snapshot config.json rather than any nested file with the same basename
• Returns None instead of raising if the config is missing, malformed, or has no valid positive integer context field
• Includes context_length in both cached-model listings and local --model listings

Tests

• New unit tests for _get_context_length cover recognized fields, priority ordering, missing fields, malformed JSON, missing files, non-positive values, string values, bool values, and None input
• New unit tests for _find_repo_config_path cover top-level-vs-nested config.json selection in Hugging Face cache snapshots
• test_handle_models now asserts every model entry includes context_length, with value either None or a positive integer

Verification

• python3 -m unittest discover -s tests -p 'test_server.py' -v
• pre-commit run --files mlx_lm/server.py tests/test_server.py
• git diff --check

Backward compatibility

Purely additive. Existing clients that only read id / object / created continue to work unchanged.