Support for DimmWitted's parallel factor graph loading

[netj]

with some compiler refactoring, e.g., moving sampler invocation details out of compiler into inference/run-sampler.

I'm still a little confused how the partitioning plays nicely with deepdive model weights reuse, but the main workflow should work as before.