Allow ROCm runners to have 2 or more gpus (#89011)
[This run](https://github.com/pytorch/pytorch/actions/runs/3432340660/jobs/5721731207) failed claiming that it couldn't detect GPUs on the runner. Inspecting the rocminfo output (higher up in logs) show that it in fact had three GPUs, but the workflow is currently setup to expect either 2 or 4 gpus.
The workflow files currently have no way of specifying wither it'll get a 2 gpu or a 4 gpu machine, so really 2 is all any test can expect to get. [This old PR](https://github.com/pytorch/pytorch/pull/72142/files) shows that historically ROCm runners only had 4 gpus, then later the logic was extended to expect 2 GPU runners as well.
It's not clear how the ROCm runner ended up with 3 gpus instead of 2 or 4 (something for ROCm folks to look into) but there doesn't seem to be a good reason for ROCm workflows to fail if 3 (or 5) gpus ever show up on a machine. This PR makes the workflows resilient to ROCm having these alternate GPU counts
Also filed https://github.com/pytorch/pytorch/issues/89012 against the ROCm team to explore why the runner only had 3 gpus
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89011
Approved by: https://github.com/huydhn
