Added TinyImageNet together with example and test.

[AndreaCossu]

1. Class labels are in the range(0, 199), so the user should use an appropriate output layer or convert labels into a smaller range. We can easily extend the class to do that based on user input.
2. The class returns numpy array (batch, channels, W, H). I saw that ImageNet leverages pytorch transforms, but since it is the only one that does so, I preferred to be consistent with all others benchmarks.
3. The class does not shuffle examples.

I can easily fix these two points in the near future but maybe we should better define how we want to manage datasets, as in this issue.

[vlomonaco]

Yes, it's good as it is now. Let's wait for #10 to be fixed based on the suggestions of #39