Learning Structured Gaussians to Approximate Deep Ensembles