Learning Performance Maximizing Ensembles with Explainability Guarantees