Boosting Deep Ensembles with Learning Rate Tuning