Fix debug_launcher issues

[muellerzr]

Fix Debug Launcher + CUDA breakage with tests

What does this add?

This PR introduces fixes for tests that run debug_launcher. They must be performed only on the CPU, as otherwise at some point if any previous tests were ran on the GPU, CUDA has been initialized for this instance of python and as a result torch will complain and break saying that a multiprocessing error has occured.

Who is it for?

closes #402

Why is it needed?

Currently running all of the tests will crash the CUDA process, due to it trying to reinitialize CUDA. From there we had the Makefile split into 2 separate calls to pytest, but given that this breakage (and since they really only run on the CPU), these tests are now under a require_cpu tag that will only run if CUDA is not available.

Note: If these tests are run first, it will not break. But for keeping it simple for the user we are going with this methodology

What parts of the API does this impact?

User-facing:

Nothing

Internal structure:

If any test should be ran on the GPU, decorate it with:

from accelerate.test_utils import require_cpu

@require_cpu
def some_test_for_cpu_only(self):
...

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks a lot!