Ability to use more meta-learner models for stacked ensembles

[dsherry]

We currently use logistic reg for classification and linear reg for regression. I bet lasso would perform better!

[angela97lin]

FWIW, the default for scikit-learn is RidgeCV (https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.StackingRegressor.html); I had gone with the linear regressor since we don't have an impl of this in EvalML yet!

[dsherry]

Note @angela97lin did some testing of this for #2093. It looked promising.

[freddyaboulton]

Based on https://github.com/alteryx/evalml/pull/2660, I think we should use ElasticNet as the meta-learner for ensembles.

FYI @christopherbunn