Shared-unique Features and Task-aware Prioritized Sampling on Multi-task Reinforcement Learning