Don't iterate over samples in the same order every epoch in a distributed training example

Don't iterate over samples in the same order every epoch in a distributed training example #22558