Introduce a `mark_sharding` that also shards the backward

[tengyifei]

🚀 Feature

PyTorch/XLA xs.mark_sharding is an in-place operation that adds sharding annotation to an XLA tensor. However, gradients to be applied to the tensor are not annotated with sharding annotations.

Motivation

In some cases, GSPMD fails to propagate sharding annotation from the tensor to its gradient. It's useful to shard both tensor and its gradient with the same sharding annotation.

Pitch

We could write a torch.autograd.Function implementation to do this.

Additional context

JAX mark_sharding shards the gradients too.

cc @bhavya01

[miladm]

cc @lsy323 for viz as he is working on disabling_spmd() API. It seems the two solutions are looking at a new subset of live tensors.