improve doc on transformation
I notice that the doc on transformation is a bit light. I think we should list there the different transformations possible, and how we should write them (there's a link to the api, but the case of the example is different (eg. crop and Crop)).
Also I saw in the tests, that the possibility to split the transforms according to the dataset split (train/validation/test) seems implemented. If it's the case, it might be good to document it.