CUDA Driver outdated in windows CI runner

[tinglvv]

🐛 Describe the bug

When pytorch CI added support for CUDA 13.0, at the very least the windows pytorch vision CI started succeeding vacuously because it would skip CUDA based testing if torch.cuda.is_available()is false. It's false because CUDA 12.2 driver's minor compatibility doesn't work with CUDA 13.0.

C:\actions-runner\_work\_temp\conda_environment_22937988283\lib\site-packages\torch\cuda\__init__.py:180: UserWarning: cudaGetDeviceCount() returned cudaErrorNotSupported, likely using older driver or on CPU machine (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\c10\cuda\CUDAFunctions.cpp:88.)https://github.com/pytorch/vision/actions/runs/22937988283/job/66591607333

pytorch/vision is not picking up the right driver, and picking up CUDA 12.2 driver: https://github.com/pytorch/vision/actions/runs/22937988283/job/66591607333

Versions

n/a

cc @atalman

[tinglvv]

The Windows test runner AMI has an outdated NVIDIA driver that still uses CUDA 12.2. The driver on that machine needs to be updated to at least 570+.

[tinglvv]

pytorch/test-infra#7826 should resolve this