* feat: bump safetensors to `0.8.0-rc.0`

* feat: run `make deps_table_update`

---------

Co-authored-by: Sayak Paul <[email protected]>

Dependency hygiene + upgrade pass:

- diffusers: drop the git source and pin >=0.38.0. The 0.38.0 release now
  contains the previously-locked dev commit (+113 commits), so the git pin is
  no longer needed.
- safetensors: override to >=0.7.0. diffusers 0.38.0 raised its safetensors
  floor to >=0.8.0-rc.0 in a metadata-only commit (huggingface/diffusers#13470,
  no code change), so pin back to the stable 0.7.x line instead of pulling a
  pre-release.
- pytorch index cu128 -> cu130: cu128 has no torch 2.12 wheels. Now resolves
  torch 2.12.0+cu130 / torchvision 0.27.0+cu130 on linux, aligned with darwin.
- raise torch floor to >=2.11.0 (torchvision >=0.26.0 to match): 2.11 is the
  first release pinning nccl>=2.28.9, which carries the pytorch#162057 fix;
  <=2.10 pin the buggy nccl 2.27.5.
- drop the nvidia-nccl-cu12 override: redundant now that the torch floor
  guarantees a fixed nccl (and cu130 ships nccl-cu13 anyway).
- remove torchaudio: unused in the tree, and cu130 has no 2.12 build.
- vendor clip-benchmark's zero_shot_classifier / run_classification into
  flow_control/third_party/clip_benchmark_zeroshot.py (GenEval colour
  classifier) and drop clip-benchmark, removing scikit-learn (its <2 cap),
  pycocoevalcap, webdataset, and friends.
- bump uv_build backend cap to <0.12.0 to match the installed uv 0.11.x.
- uv lock --upgrade: transformers 5.9, peft 0.19.1, pydantic 2.13.4, rich 15,
  starlette 1.2, datasets 4.8.5, and other transitive bumps.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>