Remove model_state.use_fp8_ddp and optimizer.all_reduce_grads

[wkcn]

Description The argument model_state.use_fp8_ddp is deprecated. In MS-AMP examples, all of model_state.use_fp8_ddp are set to True. Besides, the function optimizer.all_reduce_grads has not been used.

Major Revision

• Remove model_state.use_fp8_ddp
• Remove optimizer.all_reduce_grads
• Remove the related unittests
• Update the unittest test_fp8linear_backward since the type of weight gradient is torch.Tensor when model_state.use_fp8_ddp is True.

[tocean]

In MS-AMP-Examples, we used optimizer.all_reduce_grads. We need to remove it from examples.