State Regularized Policy Optimization on Data with Dynamics Shift