look into ensembling top model predictions

[ardunn]

Top models can be stored in the pipeline object, then can be used to make predictions. Which top models are retained should be constrained relative to how the actual top model is doing; i.e., if top model #3 is within ~95% validation error of top model #1, retain it for ensembling