Add an example with data distribution instead of epoch distribution
Background and Motivation
All official examples for Horovod appear to illustrate how epochs can be distributed across multiple MPI ranks, where each rank uses the entire dataset for training. This approach to distributed training is not always useful, as distribution of training and validation data across the MPI ranks is what's required when working with large datasets. Therefore, an example that shows how to distribute training/validation data instead of epochs can be very useful for the audience of this tutorial.
Proposed solution
Prepare and include an example that illustrates how data distribution can be easily achieved by using the slns.distribution
module.