[BE] Remove fast_nvcc tool (#96665)
As of CUDA-11.4+ this functionality can be mimicked by passing
[`--threads`](https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/#threads-number-t) option to CUDA compiler
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96665
Approved by: https://github.com/atalman, https://github.com/PaliC
