Add the image extra: transformers, torchvision and pillow

[alanprior]

Layer 1 of 4 in stack #402: the image optional-dependency group. #401 adds the encoder internals, #400 the transformer and estimator wrapper, #398 the example and docs. Only pyproject.toml and uv.lock change here, so the dependency footprint can be reviewed on its own.

What it installs

pip install "tabpfn-extensions[image]" adds transformers>=4.57.1, which loads the DINOv3 encoder and its image processor, torchvision>=0.20, which that processor is built on, and pillow>=10.4, which decodes the images. The extra joins the all meta-extra. The lock gains transformers, tokenizers, regex and torchvision; nothing is removed.

Design decisions

• The extension uses the model's own image processor from transformers, the way MulTaBench does, so torchvision is required: transformers ships no PIL backend for the DINOv3 processor. torchvision pins torch exactly; under lowest-direct that is torch 2.5.0, which tabpfn's own floor already selects.
• transformers 4.57.1 is the floor: 4.57.0, the first release with DINOv3, is yanked on PyPI.
• The added packages are Apache, MIT or BSD licensed and pass the repository's license check.

🤖 Generated with Claude Code