Sample-Efficient and Safe Deep Reinforcement Learning via Reset Deep Ensemble Agents