mlx-lm adds native Qwen3.5 MTP

This is a real local-inference win, not just a lab trick, because the speedup comes from using Qwen3.5’s built-in MTP head rather than bolting on a separate draft model. The caveat is that the feature is optimized for single-stream generation, so batching and some model variants may not benefit as much.

• –Benchmarks on an M4 Pro show about a 1.5x throughput lift, which is the kind of improvement users can feel immediately.
• –MTP disables batch serving, so this is best for interactive chats and solo generation rather than high-concurrency workloads.
• –Users may need to reconvert checkpoints to preserve `mtp.*` weights before the new path works correctly.
• –The PR discussion suggests dense Qwen3.5 checkpoints see better acceptance than MoE variants, so gains may vary by model size.