Migrate `torch_xla.device()` to `torch.device('xla')`

[ghpvnist]

The PyTorch world is more familiar with the latter and the two are basically equivalent. Instead of reinventing, let's reuse what everyone else is familiar with.