Store split information inside a Dataset

[mdekstrand]

To reduce disk storage, it would be useful to be able to directly store split information inside a Dataset, and then load a training dataset or a test item collection directly from one copy of the data.

This would work well with #506 and/or #1066.