train_ddp, process_group: fixes so CUDA works e2e

[d4l3k]

ProcessGroupBaby

This adds get_future() support to BabyWork which is required for using ProcessGroupBaby with the torchft ddp integration.

This uses a thread with an extra queue to handle future completions.

Notably this is partial support:

• the future returned currently is a None future rather than propagating the tensors which works fine as these are inplace operations
• only one of wait/get_future can be called, if you try to call both it will throw an error

ProcessGroupBabyNCCL

Fixes so it actually uses NCCL instead of Gloo and updated unit test to pass.

Manager

Fixes so we always apply the state_dict from the main thread in a safe spot to avoid version counter errors.

Test plan:

pytest

run train_ddp.py on two GPUs