Write method to register dataset

[einbandi]

Think about whether it matters when the dataset is registered (before or after defining the training phases).

I think that a validity check of keys for accessing batch attributes is only possible for the pre-defined loss types, which have keyword parameters to set the keys. The values of these could be checked on the dataset and wppropriate warnings/error could be created. These checks can be performed regardless of definition order before the training, by iterating once over all the pre-defined losses in the phases.