Created by: chenwhql

PR types

New features

PR changes

APIs

Describe

In order to simplify the writing of dynamic parallel training, remove DataParallel.scale_loss & apply_collective_grads.

• DataParallel.scale_losss -> loss.backward()
• DataParallel.apply_collective_grads -> optimizer.step/minimize

TODO:

• remove two api in all related api example code (in next PR)
• move scale_losss and apply_collective_grads into C++ for better preformance (in the future, now only for API stability of 2.0-RC)