fix: bump mlx-lm minimum to 0.31.0 for hybrid model batching

[krystophny]

Summary

Bump the mlx-lm minimum version from >=0.30.5 to >=0.31.0.

Why

ArraysCache gained native batching support (extract, merge, filter, prepare) in mlx-lm 0.31.0. Older versions crash with ArraysCache.__init__() missing 1 required positional argument: 'size' when continuous batching encounters hybrid models like Qwen3.5 that mix KVCache and ArraysCache layers.

The ensure_mamba_support() monkey-patch is already correctly disabled since these methods are native. The only missing piece was the version floor in pyproject.toml.

Reproduction

# With mlx-lm < 0.31.0:
vllm-mlx serve mlx-community/Qwen3.5-35B-A3B-6bit --continuous-batching True
# ERROR: ArraysCache.__init__() missing 1 required positional argument: 'size'

Verification

python -c "from mlx_lm.models.cache import ArraysCache; print(hasattr(ArraysCache, 'extract'), hasattr(ArraysCache, 'merge'))"
# True True (on 0.31.0+)

Files

• pyproject.toml

Fixes computor-org#11
Related: #160, #159

[waybarrios]

Now that #183 landed with the scheduler fixes for mlx-lm 0.31.x, this version bump is the matching piece. Bumping the floor to 0.31.0 prevents users from installing older versions that are incompatible with the current codebase (ArraysCache native batching, _make_cache 3-arg signature, prompt_checkpoints tuple).