Clean up workspace before test

[will-cromar]

Should fix this error:

Import check...
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/__w/xla/xla/torch_xla/__init__.py", line 152, in <module>
    from .version import __version__
ModuleNotFoundError: No module named 'torch_xla.version'

https://github.com/pytorch/xla/actions/runs/8898851270/job/24437801900

This results from combination of questionable decisions from both GitHub and Python. First, GHA does not guarantee that the workspace is actually empty. In this case, torch_xla/ is still checked out from a previous run. Since there is a local torch_xla directory, Python gives this precedence over the installed torch_xla package.

Just delete everything in the workspace before running tests to eliminate non-determinism.

[will-cromar]

This is a little repetitive. We can package the environment setup in a reusable template if we start adding more workflows that need it.