Compositional Visual Generation with Energy Based Models