Sticking the Landing: Simple, Lower-Variance Gradient Estimators for Variational Inference