Switch to the standard setup-nvidia GitHub action (#264)

Summary:
As all PyTorch repos are now running on share runners as part of project Nova. We need to avoid code duplication when copying `install_nvidia_utils_linux.sh` and older NVIDIA driver by switching reusable [setup-nvidia](https://github.com/pytorch/test-infra/blob/main/.github/actions/setup-nvidia/action.yml) from test-infra repo.

For the context, I'm currently seeing some NVIDIA installation failures from PyTorch where the older driver `510.60.02` is used, so I'm trying to roll out this new GitHub action to avoid this:

* https://github.com/pytorch/pytorch/actions/runs/3384224730/jobs/5621261503
* https://github.com/pytorch/pytorch/actions/runs/3383111924/jobs/5618889208
* https://github.com/pytorch/pytorch/actions/runs/3382298003/jobs/5617245371

Pull Request resolved: #264

Reviewed By: PaliC

Differential Revision: D41190057

Pulled By: huydhn

fbshipit-source-id: f0bacd00d5963ad1757a49d54a3711e7cb9da2d0

Author: huydhn

.github/scripts/install_nvidia_utils_linux.sh

@@ -20,19 +20,8 @@ jobs:
         uses: actions/checkout@v2
         with:
           submodules: true
-      - name: Clean up previous CUDA driver installations
-        shell: bash
-        run: |
-          set -x
-          yum list installed | grep nvidia || true
-          yum list installed | grep cuda || true
-          sudo yum remove -y cuda || true
-          sudo yum remove -y cuda-drivers || true
-          sudo yum remove -y "*nvidia*" || true
       - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
-        run: |
-          bash .github/scripts/install_nvidia_utils_linux.sh || true
-          echo "GPU_FLAG=--gpus all" >> "${GITHUB_ENV}"
+        uses: pytorch/test-infra/.github/actions/setup-nvidia@main
       - name: Setup SSH (Click me for login details)
         uses: ./.github/actions/setup-ssh
         with: