Stochastic Dimension-reduced Second-order Methods for Policy Optimization