`clang-tidy` workflow only needs `cuda-toolkit` (#36241)
Summary:
`cuda` metapackage install both kernel driver + runtime libraries + toolchain, while `cuda-toolkit` metapackage as name suggests installs only toolchain + library headers.
This reduces install dependencies time for `clang-tidy` step by 60+ sec
Pull Request resolved: https://github.com/pytorch/pytorch/pull/36241
Test Plan: CI
Differential Revision: D20923839
Pulled By: malfet
fbshipit-source-id: 1e773285222bed179973449573215fcaee1983de