Implement AMP

[edoyango]

AMP speeds up training pretty significantly on NVIDIA GPUs. However, it's not implemented in the dynamicdet code. Perhaps can dig around in the ultralytics code to figure out how they do it.

Bread crumbs:

• https://pytorch.org/tutorials/recipes/recipes/amp_recipe.html
• https://github.com/ultralytics/ultralytics/blob/b63af4cbcc15cbafc41f9d288b00655df1357136/ultralytics/engine/trainer.py#L57