Introduce no_sync context wrapper + clean up some more warnings for DDP

[muellerzr]

Introduce no_sync context wrapper

What does this add?

• This PR adds a no_sync wrapper that shadows torch.nn.parallel.DistributedDataParallel.no_sync but will do nothing if we're not in a DDP process
• It also refactors launch args for torchrun into a separate util to make reusing it in tests + launch easier

Why is it needed?

This lets us use Accelerator eventually for gradient accumulation

What parts of the API does this impact?

User-facing:

• Introduces Accelerator.no_sync(model) to perform the equivalent of model.no_sync() (if in ddp, if not it does nothing)

Non-user facing:

• Introduced tests that mimic the official tests for no_sync in PyTorch to ensure those work, and CPU tests to make sure that when not in DDP it's a noop context manager

[muellerzr]

@sgugger good for a rereview, slow test logs can be seen here passing: https://github.com/huggingface/accelerate/runs/6794400970?check_suite_focus=true

[sgugger]

Thanks for iterating on this!

Just missing a by-feature example of use :-)