fix: detect vision patch size for unsplit (HF-format) Qwen3-VL

[fszontagh]

Summary

HF/ComfyUI-format Qwen3-VL text encoders keep the vision patch embed unsplit as visual.patch_embed.proj.weight. The existing detection only matches the split visual.patch_embed.proj.0.weight, so vision.patch_size silently stays at the Qwen2.5-VL default of 14 instead of 16. Any model that builds the Qwen3-VL vision tower (e.g. Mage-Flow) then fails to load such an encoder with patch_embed.proj.weight got [16, 16, ...], expected [14, 14, ...].

This reads the patch size from the unsplit weight's kernel extent (ne[0]), and additionally derives hidden_size from visual.patch_embed.proj.bias (the split name visual.patch_embed.bias does not match the unsplit .proj.bias). GGUF-sourced and split-format encoders are unaffected.

The model.language_model.* naming difference is already handled by scripts/convert_qwen3_vl.py; this only fixes the remaining patch-size detection, which that script does not touch.

Related Issue / Discussion

Related to #1708 - users hit this loading the Comfy-Org bf16 Qwen3-VL; the convert script fixes naming but the unsplit patch size still mis-detects.

Additional Information

Verified with the bf16 Qwen3-VL-4B from Comfy-Org/Krea-2, run through scripts/convert_qwen3_vl.py, then loaded as --llm for Mage-Flow: before this change it fails with patch_embed.proj.weight ... expected [14, 14, ...]; after, it loads and generates correctly. GGUF encoders continue to work unchanged.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.