The Power of Learned Locally Linear Models for Nonlinear Policy Optimization