RPC: Megatron-LM support

[leezu]

Pytorch-Lightning has a great support for ZeRO via DeepSpeed and Fairscale plugins. Unfortunately ZeRO-3, runs slower than ZeRO2, it hard to keep good throuhgputs when scaling to very large models. Megatron-LM on the other hand can sustain similar throughputs as ZeRO-2 and Megatron-LM integration would be useful in Pytorch-Lightning.

I see #8101 nuked the existing RPC plugin. I'd like to start a discussion to add them back, perhaps supporting both torch 1.9 Pipe and Megatron-LM?