Reduce scale factor for DDP in generative model

[Can-Zhao]

Is your feature request related to a problem? Please describe.
Currently, scale factor is computed for each rank for DDP in generative model.
It should be reduced across ranks to use the same scale factor.
It does not cause a problem for performance for now because scale factors are all close to one in this example.
But in the future it might cause issue.

Same issue for validation loss.

Describe the solution you'd like
Reduce scale factors and validation loss across ranks.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.