Speed up incremental and CI builds

On my personal machine which is somewhat faster than our Cloud-based development VM, splitting these crates brings down the time to do cargo xtask after a trivial change to a harvester from 14s to 9s, i.e by more than a third.

This is somewhat difficult to review but the first three commits are really only shuffling files and dependencies around without any functional changes.

The fourth commit changes how expectations are stored for our regression tests which currently slows down the CI significantly as the binary datasets are stored using Git LFS. After this change, we just store the pretty-printed JSON directly using Git (without LFS) which should be faster to fetch and is easier to diff.