Solving Compositional Reinforcement Learning Problems via Task Reduction