Proper CPU allocation when submitting to multiple GPU nodes

[vyasr]

When requesting more GPUs than are available on a single node during submission, signac-flow needs to consider how the CPU tasks are distributed. It doesn't make sense to allocate all the CPU tasks onto a single node or do any math related to that. As suggested in #233, a sensible choice would be to evenly divide the CPU tasks per node, at least in cases where all jobs acquire full nodes. The question is more complicated in cases where there are shared nodes; we may just have to default to the dumb assumption and require users to provide their own templates to do more customized calculations. In that case, we probably should document more thoroughly what choices we've made in the signac-flow templates documentation.