Remove non-parallel approaches to get a device

[yaoshiang]

🐛 Bug

There are four ways to get a device object. Three could be deprecated as they are likely unnecessary, and also anti-parallel to torch.cuda.device(), which does not return a device but is in fact a context manager.

torch.device('xla', 3)
torch_xla.device(3)
torch_xla.torch_xla.device(3)
torch_xla.core.xla_model.xla_device(3)

To Reproduce

Run this in a collab with a TPU enabled and pip installed pytorch_xla with the correct version.

torch.device('xla', 3)
torch_xla.device(3)
torch_xla.torch_xla.device(3)
torch_xla.core.xla_model.xla_device(3)

Expected behavior

Should deprecate the three ways to get a device inside torch_xla.
Docs that use them should be updated to torch.device("xla", <optional int>) and remove import of torch_xla.core.xla_model if necessary.

Environment

TPU on Collab

Additional context

Should verify that the four APIs do ultimately call the same underlying function to guarantee equivalent behavior.

[ysiraichi]

I agree. We should not have this many ways to get an XLA device. That said, I'm not against having one internal torch_xla.device(3) function, so that we have control over that.

cc @miladm @zpcore @tengyifei @ManfeiBai

[miladm]

Awesome.

@yaoshiang do you think we can set up a deprecation API github issue as the central source of trust for what we intend to do on this line of work / roadmap?