Exponential Tilting of Generative Models: Improving Sample Quality by Training and Sampling from Latent Energy