Efficient Multi-Task and Transfer Reinforcement Learning with Parameter-Compositional Framework