Add Qwen3-VL dense models

[pengdurice]

Add support for Qwen3-VL dense models, locally tested with Qwen3-VL-4B-Instruct, Qwen3-VL-8B-Thinking, Qwen3-VL-32B-Instruct.

---

Note

Medium Risk
Adds a new architecture JSON that affects how MergeKit identifies and maps Qwen3-VL dense model weights; incorrect keys/weight names could break loading or merges for these models. Change is isolated to data/config with no broader code logic modifications.

Overview
Adds a new modular architecture definition qwen3_vl_dense.json to support dense Qwen3-VL models (Qwen3VLForConditionalGeneration).

The definition introduces module weight mappings for the text decoder (dense MLP projections), the multi-modal projector (deepstack merger layers), and the vision tower blocks, along with updated config keys for vision_config.depth and text_config.vocab_size and required tagalong files.

^{Written by Cursor Bugbot for commit f1f257f. This will update automatically on new commits. Configure here.}

[cg123]

Thanks for the contribution!