Add qwen3_vl_moe to VLM registry

[JannikSt]

• Trainer crashes with Unrecognized configuration class Qwen3VLMoeConfig for AutoModelForCausalLM when loading Qwen3-VL-30B-A3B-Thinking
• Missing qwen3_vl_moe entry in VLM_REGISTRY, so VLM detection fails and it falls through to AutoModelForCausalLM instead of AutoModelForImageTextToText
• One-line fix: add the model_type to the registry with the same vision/language attrs as the other Qwen VLMs

---

Note

Low Risk
Low risk, single-entry registry update that only affects model-type detection and component attribute resolution for qwen3_vl_moe. Potential impact is limited to runs loading that specific model family.

Overview
Adds qwen3_vl_moe to VLM_REGISTRY in vlm.py, mapping it to the same vision_encoder_attr/language_model_attr as other Qwen VLM variants.

This ensures configs with model_type="qwen3_vl_moe" are detected as VLMs and use the VLM-specific component resolution paths instead of falling back to text-only handling.

^{Reviewed by Cursor Bugbot for commit 3d45fe9. Bugbot is set up for automated code reviews on this repo. Configure here.}