feat: support DynamicLossScale for TrainStep

[geniuspatrick]

Thank you for your contribution to the MindCV repo. Before submitting this PR, please make sure:

• [x] You have read the Contributing Guidelines on pull requests
• [x] Your code builds clean without any errors or warnings
• [x] You are using approved terminology
• [ ] You have added unit tests

Motivation

• GradientAccumulation inherits from mindspore.boost.GradientAccumulation
• Support all combinations of loss_scale_type and drop_overflow_update.
  1. loss_scale_type="fixed", drop_overflow_update=False --> update_cell=None, TrainStep=TrainOneStepCell(scale_sense=loss_scale)
  2. loss_scale_type: fixed, drop_overflow_update: True --> update_cell=FixedLossScaleUpdateCell, TrainStep=TrainOneStepWithLossScaleCell(scale_sense=update_cell)
  3. loss_scale_type: dynamic, drop_overflow_update: True --> update_cell=DynamicLossScaleUpdateCell, TrainStep=TrainOneStepWithLossScaleCell(scale_sense=update_cell)

Test Plan

(How should this PR be tested? Do you require special setup to run the test or repro the fixed bug?)

Related Issues and PRs

256

469

483

489

569

609

fixes #604