Fix PyTorch master break - use proper THTensor_storage() API

Bugfix mpi4py: Create a private MPI communicator

Collective control plane & other low latency improvements

Bugfix: `python setup.py install` requires cffi

Support TensorFlow Optimizers that override compute_gradients/apply_g…

…radients

PyTorch support for Horovod

Avoid deadlock if worker crashes with an exception

New major release

Add hvd.local_size() API

Add Keras ImageNet training example and LearningRateScheduleCallback