fix(rerank): pin torch to cu128 wheels for wider CUDA compatibility

[shan-nvidia]

Problem

The rerank eval stage (stage3_eval) resolved PyTorch from PyPI, which
pulls wheels compiled against the latest CUDA toolkit (currently 13.0).
This breaks on any server whose NVIDIA driver does not yet support
CUDA 13.0, for example:

CUDA initialization: The NVIDIA driver on your system is too old (found version 12090).

Fix

Pin torch to the cu128 (CUDA 12.8) wheel index, which is
forward-compatible with CUDA 12.8, 12.9, 13.0, and newer drivers.

• Added torch as a direct dependency (previously only a transitive dep
  via sentence-transformers).
• Configured a [[tool.uv.index]] pointing to
  https://download.pytorch.org/whl/cu128 with explicit = true, so
  only torch resolves from it — all other packages continue to use
  PyPI.
• Added [tool.uv.sources] mapping torch to the cu128 index.
• Regenerated uv.lock.

Files changed

File	Change
src/nemotron/recipes/rerank/stage3_eval/pyproject.toml	torch cu128 index + direct dep
src/nemotron/recipes/rerank/stage3_eval/uv.lock	Regenerated

Test plan

• Run nemotron rerank eval on a CUDA 12.x driver — verify torch initializes CUDA successfully

[shan-nvidia]

Close this since it's already fixed.