Batching should handle outputs too

[robbason]

e.g. when doing a pose prediction challenge, the protein may be put into a separate reference frame by the container's prep. So we could ask the container to return the protein structure as well. But if running hundreds of ligands, we'd end up with hundreds of duplicate protein structures.