mnist example contains bad practices leading to wrongful results

In the mnist example there should be explicit comments about the default nature of the `F.nll_loss` which is to average the results over the mini-batch. The other thing that can lead to errors is the `len(train_loader.dataset)` even though that returns the number of samples in the `train dataset` if someone was to create a split of train data into train/valid using `SubsetRandomSampler` then calling again `len(train_loader.dataset)` will still return you the full number of samples in the train set (which is wrong) without reflecting the correct number of samples since the train set now contains X% less samples which have been kept as validation set.