Sequential Action-Induced Invariant Representation for Reinforcement Learning