Sample-efficient Model-based Reinforcement Learning for Quantum Control