Safe Policy Optimization with Local Generalized Linear Function Approximations