ENH refactor the way normalization is being applied to the datasets

Currently, there is a lot of copy-pasting that was introduced by https://github.com/benchopt/benchmark_resnet_classif/pull/19 when it comes to handling the normalization.

Basically, we want to have the same normalization for all datasets, but not apply it at the same times.
In particular, we want to be able to apply it after data augmentation in the case of the training set when fitting the model.

Start by reading pull request #19 and tracing how each dataset currently applies normalization and data augmentation. Compare the training and evaluation paths; done means all datasets share the normalization logic while training applies it after augmentation.

Written by the indexing model from the issue text.