GitLab

* add torch.cuda.amp to fp16_utils and optimizers

* use with context manager for autocast

* add doc to explain the behavior differences between real amp and ours

* fix docstring