Upgrade NVTX to NVTX3 (#90689)
Due to recent upgrade to CUDA 11, we can upgrade NVTX to NVTX3 as well, which is a header only library that can simplify the building system a lot.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90689
Approved by: https://github.com/soumith, https://github.com/malfet
