Learning Convex Regularizers for Optimal Bayesian Denoising