Small data support for ensembles: use out-of-sample preds to train metalearner

[dsherry]

Problem
In the case when we don't have many samples (<1k total) provided to automl, we should adjust our splitting strategy to maximize our usage of what data we have available.

Background
#1746 tracks creating a separate 20% split for stacked ensemblers.

Proposal
If our data is "small", we should instead do the following:

• Do not create a separate split for stacked ensemblers
• Perform normal CV during pipeline training, as we do now
• When training stacked ensemble method, use the complete set of each pipelines' predictions on each validation split during CV to train the metalearner.

We still need to resolve: what data do we use to validate the ensembler?