Explain on the fly transforms for PyTorch

[adam-narozniak]

Issue

So far, immediate transforms have been used to create PyTorch DataLoader from datasets.Dataset. It is not how the typically downloaded Dataset from PyTorch torchvision behaves. The transforms applied to it execute at the access time. The difference might bring confusion to users or introduce unexpected behavior.

Proposal

Explain how to do the transforms on-the-fly that strictly resembles the PyTorch way of doing it (which users are probably familiar with and expect).
The instant transform to all the dataset still exist in the documentation but the difference in the behavior is explained. Also, it's explained after the on-the-fly transforms.