Scaling Up Bayesian Uncertainty Quantification for Inverse Problems using Deep Neural Networks