Better checks for if a TPU device exists

[muellerzr]

This PR fixes a situation where torch_xla is installed but a TPU device isn't present.

This originally would cause the import to pass, even though a TPU device isn't found. Instead we combine this with a quick xm.xla_device(), that raises a RuntimeError if not on a TPU.

I've tested this on a CPU and TPU version of Colab.

cc @sgugger

[HuggingFaceDocBuilderDev]

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

[sgugger]

Thanks a lot!