Pipeline parallelism requires special gradient clipping implementation

[le1nux]

The current implementation of the gradient clipper only works with DP, FSDP and HSDP.
For PP, the gradients of the different stages would not be accumulated.

See:
https://pytorch.org/docs/stable/fsdp.html#torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_

https://github.com/pytorch/torchtitan/blob/b291ad662493b63d25b038a30a915082d3617baf/torchtitan/distributed/utils.py#L256-L259