Towards Versatile Humanoid Table Tennis: Unified Reinforcement Learning with Prediction Augmentation