Ensembling: provide original input data as features to metalearner

**Background**
Currently our ensembler components are provided as features only the predictions or predicted probabilities from the input pipelines.

**Proposal**
Alongside the input pipeline predictions or predicted probabilities, include a preprocessing chain which additionally provides the processed input features to the metalearner.

Our new automl algo first constructs a preprocessing chain, then applies that same chain to each estimator. We should use the same chain as input to the metalearner.

Filed as a spike because we should agree on the design before merging any code which changes the behavior of automl.