Learning an Astronomical Catalog of the Visible Universe through Scalable Bayesian Inference