Avoid introducing unnecessary system dependency on CUDA (#15449)
## Summary
Packages like `triton` should come from the PyTorch index, but they
don't actually vary across (e.g.) the `cu128` or `cu129` indexes.
Closes https://github.com/astral-sh/uv/issues/15446.
## Test Plan
Validate that the following pins to `cu128`, rather than `cpu`:
```
echo "vllm\ntorch==2.7.1+cu128" | cargo run pip compile --torch-backend=auto --extra-index-url https://wheels.vllm.ai/b2f6c247a9b84556a8ea0e75bb4a2db765ff3315 - --python-platform linux --python-version 3.13 -v
```
