Model-based Reinforcement Learning with Probabilistic Ensemble Terminal Critics for Data-efficient Control Applications