No Prior Mask: Eliminate Redundant Action for Deep Reinforcement Learning