Force allreduce of all gradients in step(), bugfixes

FP16 hierarchical allreduce

PyTorch 1.0, TF-Keras, FP16 ops on GPU

Parallelized hierarchical allreduce

Support for the upcoming PyTorch release

Add compatibility with PyTorch 0.4.1

Support for IBM PowerAI DDL & APIs to restore optimizer state

Critical Bugfix: PyTorch must wait for GPU data before allreduce

Critical Bugfix: non-fused allreduce produces incorrect results

Hierarchical allreduce & differentiable ops