reinstate apex.amp (O1 O2)

[vince62s]

For benchmark and better documentation on performance I am adding back Apex.amp.

As a reminder, we have:
For adam:
if fp32, then it will use the native fp32 adam training.
if fp16, then it will use the "amp" library included in pytorch.

For fusedadam:
if fp16 or fp32 and "opt.apex_opt_level" is NOT set, then it will use the old legacy FusedAdam algorythm from the original apex implementation. https://github.com/NVIDIA/apex/blob/master/apex/contrib/optimizers/fp16_optimizer.py

if "opt.apex_opt_level" is set to:
"O0": training will be done in fp32 whatever the opt.model_dtype flag.
"O1" or "O2" will train according to those parameters: https://github.com/NVIDIA/apex/blob/master/docs/source/amp.rst
"O3": does not work with fusedadam, so do not use.