fix: pin torchvision to the cu126 index so the vision extra imports

[rasros]

uv sync --extra vision resolved torchvision from PyPI while torch came from the pinned pytorch-cu126 index. The mismatched builds import fine individually but blow up when transformers touches torchvision ops:

RuntimeError: operator torchvision::nms does not exist

That surfaced as the vision scorer silently falling back to SimpleFallbackScorer under --scorer auto, since the AutoImageProcessor load failure is caught and downgraded.

Two changes:

• Declare torchvision explicitly in the vision extra — it was only ever an implicit transitive need of the SigLIP image processor.
• Pin it to pytorch-cu126 in [tool.uv.sources], matching torch.

Verified after the change: torch 2.13.0+cu126 | torchvision 0.28.0+cu126, SigLIP processor loads, and a run reports Using vision scorer instead of falling back.