Rethink the way splits are saved

[xroynard]

There are 2 kinds of splits :

• those defined by the dataset creator (should they be stored with the Dataset._infos ?), including:
  • "canonical" splits, they form partition of the ensemble
  • "secondary" splits
• those defined when training a model:
  • they should be stored in the ProblemDefinition ?
  • there names should be standardized : train, val, test ?
  • they should not overlap ?
  • they can make references to splits defined in Dataset and/or to sample ids