[REQUEST] Pipeline Parallelism support multi optimizer to train

[whcjb]

Is your feature request related to a problem? Please describe.
i want to train big model gan, must shard to multi-gpu, but the pipeline_module seems not support multi-optimizer
Describe the solution you'd like
support

Describe alternatives you've considered
can control flow to which layer

Additional context