change the documentation

[stephandooper]

Chnage the documentation to reflect the following changes:

Best practices:

• Use of mixed precision -> should be float-16-true, not mixed
• Using metric collections if you want to try standard image classification models to track metrics, prevents unnecessary subclassing
• copy to gpu -> implement a transfer_batch_to_device hook if you want the False option
• Using normalize_on_gpu: advice to set to True
• Using memory format channels last

Custom models:

• Use of the streamingconstructor for making custom models, instead of current workflow where nothing is done