fix(qwen): remap legacy single-file VL encoder keys under transformers 5.x

[Pfannkuchensack]

Summary

Qwen2_5_VLForConditionalGeneration._checkpoint_conversion_mapping is empty in transformers 5.x, so the if key_mapping: guard skipped remapping and ComfyUI single-file checkpoints (visual.*, model.*) failed to load: their keys were reported as unexpected while model.visual.* params stayed as meta tensors. Fall back to the legacy mapping when transformers no longer provides one. The negative lookahead keeps new-layout checkpoints untouched.

Related Issues / Discussions

N/A

QA Instructions

1. Obtain a ComfyUI-style single-file Qwen2.5-VL encoder checkpoint (e.g. qwen_2.5_vl_7b_fp8_scaled.safetensors from the starter models) that uses the legacy key layout (visual.*, model.layers.*).
2. Install it as a Qwen VL encoder and run a qwen_image_text_encoder invocation.
3. Before this fix: load fails with RuntimeError: Failed to load all parameters from checkpoint. Meta tensors remain: ['model.visual.patch_embed.proj.weight', ...] plus a 728 unexpected keys warning.
4. After this fix: the encoder loads cleanly with no unexpected-key warning and no meta tensors remaining; text encoding completes normally.

Verified against the real checkpoint header (1446 keys): 730 → model.language_model.*, 714 → model.visual.*, 1 → lm_head.weight, 1 → scaled_fp8 (fp8 metadata, stripped before remap). No weight keys left unmapped.

Merge Plan

Standard merge — no DB/schema changes, no migrations.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[lstein]

Works as advertised.

[lstein]

Supersedes #9334