Wide performance discrepancy between LM Studio MLX and `mlx_vlm.generate` with Qwen3.5-35B-A3B

When running `Qwen3.5-35B-A3B` [4bit MLX](https://huggingface.co/mlx-community/Qwen3.5-35B-A3B-4bit) on an M4, I noticed that using a ~3000 tokens prompt:

- LM Studio gets consistently around **78 tok/s** token generation and **6s TTFT**
- `mlx_vlm.generate` reaches **108 tok/s** token generation and **2s TTFT**

The discrepancy in token generation is almost identical even with very short prompts (<100 tokens).

Start by reproducing the Qwen3.5-35B-A3B 4bit benchmark on an M4 with the reported ~3000-token and short prompts, comparing LM Studio against mlx_vlm.generate. Investigate the generation-speed and time-to-first-token difference; done means identifying and documenting the cause or producing a focused fix with measurements showing the discrepancy is resolved.

Written by the indexing model from the issue text.