Apex support for mixed precision training

[BloodAxe]

Makes mixed precision training as simple as possible:

model, optimizer = amp.initialize(model, optimizer, opt_level="O1")

runner = SupervisedRunner()
runner.train(
    model=model,
    criterion=criterion,
    optimizer=optimizer,
    scheduler=scheduler,
    ...)

Breaking changes (mostly internal, but still):

1. Experiment.get_optimizer renamed to Experiment.get_optimizer_and_model and returns optimizer and model as a tuple. Motivation: Apex requires two objects to initialize - a model and optimizer. A model must be already on GPU and not wrapped with DataParallel/DistributedDataParallel. This change was made to have clear amp initialization in one place.
2. UtilsFactory.prepare_model now does not wrap model with DataParallel since this will break Apex initialization. Instead, DataParallel applied in Experiment.get_optimizer_and_model if fp16 is not enabled and more than one GPU is available.

[Scitator]

@BloodAxe proposal: add fp16 flag to SupervisedRunner .train and .infer with automatic model cast to fp16 mode. What do you think about it?

[BloodAxe]

This need discussion. Apart from global enable/disable fp16 training mode flag, there are couple of flags that user may want to change (optimization level, flag to keep batch-norm in fp32, loss scaling flags, etc).
This would expand set of parameters in SupervisedRunner. That's my only objection so far.

[Scitator]

@BloodAxe hmm, you are right...
then, what about fp16 flag and the following logic:

• do nothing if model & optimizer are already fp16
• turn default optimization level (01?), keep batch norm in fp32 and loss autoscaling (or scale by 128)
  ?

[BloodAxe]

@Scitator I think it's still make sense to have fast&quick option to turn on fp16 mode for SupervisedRunner.
I'll fix remaining issues and add fp16 flag in a while.