Add NVIDIA GPU runners and run CUDA tests again

[ScottTodd]

We used to have some Linux runners with NVIDIA T4 and A100 GPUs on the GCP runner cluster. Our new Azure runner cluster currently only has Linux CPU runners.

We have unit tests and larger test suites using the CUDA and Vulkan HAL. None of these are particularly CPU heavy, so we could get by with a 4 or 8 core CPU and an attached GPU, if such a configuration is available.

Target should be presubmit (pull_request and push events), with a load of around 50-100 (up to 400) runs per day.

Can start with a single GPU type, but eventually we could also run tests across a wide range of data center and consumer cards nightly, like T4, A100, H100, 1080, 2080, etc. We will also want to run benchmarks eventually, which will need persistent runners with cached model weights and some tuned hardware/driver settings.