[qwen3 vl] [qwen2 vl] Add multimodal RoPE and Deepstack implementation to mlx-lm language model port

[neilmehta24]

Qwen3 VL models rely on "Deepstack" embeddings to improve image-to-text capabilities. This is not implemented in mlx-lm yet, so OCR capability with the unified arch (i.e. vision_add_on) is subpar compared to the mlx-vlm implementation. Ref: 1 2

[neilmehta24]

Also seeing the same issues with qwen2_5_vl and qwen2_vl. My theory is that the multimodal RoPE implementation is critical to these models working correctly.