Core functionality to allow multi-device workflow

[orlando-labs]

Hi @ankane, as we discussed before, I’ve kept only the core functionality needed for the multi-device workflow. All of the distributed features have been moved to a separate repo/gem.

[ankane]

Hi @orlando-labs, thanks for another PR. This still has more changes than I'd like to include.

I've fixed the item method in the commit above and addressed the TODO for changing Tensor#device to return a Device.

I'd be happy to include a minimal PR for adding map_location and weights_only options to Torch.load if it closely follows the Python code, but I think the rest should probably live in the separate repo.

[ankane]

Two other things that I think could be included (in two separate minimal PRs) would be Torch::CUDA.set_device and the Torch::Accelerator module.