Add better cross-validation behavior.

[mthomure]

When cross-validation is used from GLAB, prototypes should optionally get re-learned for each sub-split. Currently, the prototypes are learned once, and then evaluation happens.

To support this, we'd need to refactor the ExperimentData object to combine the extractor and evaluation data. Alternatively, we could have cross-validation return multiple ExperimentData objects, where each corresponds to a different sub-split.