Reinforcement Learning with Delayed, Composite, and Partially Anonymous Reward