Improve testing of datasets

[pmeier]

One reason we lag behind with the datasets tests (see #963 (comment)) is that right now writing a test is lengthy and complicated. To overcome this, I'm proposing to add a well-documented DatasetTestCase superclass. With minimal user interference it should be able to automatically test

1. if the dataset inherits from VisionDataset,
2. if the dataset raises an error in case it does not exist or is corrupted,
3. if applicable test if the transform, target_transform, and transforms are applied properly,
4. if repr(dataset) can be built, and
5. if len(dataset) matches the number of fake examples (see below).

Of course the contributor is free to test more things that go beyond the above list, but in general this should cover most cases for most datasets. The only thing that needs to implement manually is the generation of fake data in the respective structure of the dataset. This can be aided by providing utility functions to create an image, a folder of images and so on.

Within the test we then simply mock out the download and extraction functions and thus make the dataset accept the fake data. Given that this logic is rarely in excess of 10 LOC and uses mostly functions from torchvision.datasets.utils, IMO it is a good tradeoff towards simpler tests.

cc @pmeier

[fmassa]

I think this is great!

One thing we should always enforce for testing as well is that __getitem__ works and gives the expected result for mocked data.

[pmeier]

Closed in #3402