Add CPU tests for aarch64 GPU wheels

[NicolasHug]

Yes you read that right: CPU tests for GPU wheels. We can build the ARM GPU wheels but we can't test them (on GPU).

[NicolasHug]

I'll remove those before merging. That's just to avoid triggering these other jobs as I was testing the new ones.

[Dan-Flores]

This cuda version is specified outside of the matrix, as done in linux_cuda_wheel. Is that intentional?

torchcodec/.github/workflows/linux_cuda_wheel.yaml

Lines 63 to 72 in 957a150

	matrix:
	# 3.10 corresponds to the minimum python version for which we build
	# the wheel unless the label cliflow/binaries/all is present in the
	# PR.
	# For the actual release we should add that label and change this to
	# include more python versions.
	python-version: ['3.10']
	# We test against 12.6 and 13.0 to avoid having too big of a CI matrix,
	# but for releases we should add 12.8.
	cuda-version: ['12.6', '13.0']

[NicolasHug]

Good catch - yes that's intentional. In this new job we only test on a CUDA 12.6 machine, not a any other CUDA versions (see image: pytorch/manylinuxaarch64-builder:cuda12.6) above.

We don't need to test more CUDA versions because we can't run CUDA tests on ARM anyway (that's a test-infra limitation, from what I understand). And we still have to use a CUDA docker image because the wheels were built with CUDA support.