Slow performance issue with the Pytorch Horovod example

In the `DataLoader` instantiation step, you need to pass `num_workers > 0` (4 caused a 2x-speedup for me on a 4 core node) to have a performance higher than the single node execution. Without this parameter passed to the `DataLoader` constructor, the performance is slower than the single `MNIST` example.

Link to the file: [https://github.com/Azure/MachineLearningNotebooks/blob/master/how-to-use-azureml/ml-frameworks/pytorch/distributed-pytorch-with-horovod/pytorch_horovod_mnist.py](https://github.com/Azure/MachineLearningNotebooks/blob/master/how-to-use-azureml/ml-frameworks/pytorch/distributed-pytorch-with-horovod/pytorch_horovod_mnist.py)

Line: 62 and 72.