Add colqwen3 embed

[bulutyigit]

PR Description

What
Adds ColQwen3 (Tomoro tomoro-colqwen3-embed-8b) support to mlx-vlm for multimodal retrieval embeddings.

Why
Tomoro ColQwen3 is a ColBERT-style multi-vector embedding model on top of a Qwen3-VL backbone.
Without native support, MLX users cannot convert/load the checkpoint or produce token-level embeddings.

Changes

• Adds colqwen3 model type integration for convert/load
• Implements embedding_proj_layer and encode() to output token-level embeddings [B,T,D]
• Adds helper APIs:
  • encode_queries(processor, texts)
  • encode_images(processor, images) (returns visual-token embeddings; ideal for PDF patches)
  • maxsim(q, d) for ColBERT MaxSim scoring
• Adds weight-key sanitization to map Tomoro/HF keys to MLX module names (vlm.model.* → vlm.*)
• Fixes hidden forward path to correctly respect masks (embedding extraction path)

Testing

• Verified load(<HF repo>, trust_remote_code=True) works
• Verified text embeddings + image embeddings produce valid shapes and MaxSim scores

Notes
This PR focuses on embedding usage, not generation. No changes to public generation APIs expected.

[Blaizzy]

Hey @bulutyigit,
Happy new year, this is an awesome addition!
I actually built a package called mlx-embeddings which would be the perfect home for this port. Would you mind redirecting the PR there? I'll review and merge it there.
Thanks for the contribution!