FSDP with FP8 is not working

[eljandoubi]

layernorm needs to have reshape instead of view.
In pytorch base module, torch.get_autocast_gpu_dtype() is deprecated.
@timmoon10 @ksivaman @ptrendx @cyanguwa

[timmoon10]

TE does not handle non-contiguous data very well. I like the idea of coercing data to be contiguous, but that adds non-trivial CPU overhead even if it has no effect on GPU.

Related: #417