Default parallelization when creating datasets

[hubertjb]

As mentioned in #288, we could uniformize the Dataset classes API a bit more, which would allow us to provide parallel loading to all datasets by default. For instance, we could require every custom dataset class implements a _create_dataset method (sometimes called _load_raw in existing datasets I think) and take care of the joblib loop by default in a create_datasets method that all datasets inherit.

@gemeinl