if somebody needs non-default behaviour, we should do this in another way!

used to split dss object in train valid and test according to event number
first map prepare:
  1. duplicates data ([012]->[01201])
  2. saves correct chunk starts and stops:
    split 0: test=[0], valid=[1], train=[2]
    split 1: test=[1], valid=[2], train=[0]
    split 2: test=[2], valid=[0], train=[1]
then map apply