process_group: register via public API

[d4l3k]

This registers the ProcessGroups via the public APIs (Backend.register_backend and dist.new_group) rather than modifying internal state.

This makes things a bit messier but fits better into the usual PyTorch init logic. This requires a global PG to be initialized. A future PR could detect if a PG is initialized and register this as the global PG if necessary.

Weirdly -- with how this works it ends up returning the exact same ProcessGroup object from dist.new_group. I'm not quite sure how that actually works. The Backend wrapper seems to unwrap itself when casted to PG.

Test plan:

pytest torchft/process_group_test.py

[fegin]

I like this approach even though it requires the global PG.