Overview

This PR adds a "narrowing" step to diff/merge. Previously, we'd compute the syntactic hash of every type and term in a namespace, then perform a diff on those hashes. Now, we avoid computing the syntactic hash of terms and types that we know (by their Unison hash) are equal. This allows us to skip a lot of work for diffs/merges that don't transitively touch a lot of definitions, such as those that contain only new definitions.

Additionally, this PR and speeds up definition loading in general by utilizing branch/names caches that were added recently in update work. Previously, we'd load all branches involved from scratch (and that's where the conflicted names check was implemented). But this was always slower than going the other direction – using the pre-built Names in the Branch object as a starting point, and checking it for conflicted names. Since update now does something similar, I deleted all of the old slow loading code and reworked the starting point of diff/merge to basically be "given branches that are not conflicted, ...".

The merge-as-a-library API was simplified down to just two steps, diff (producing a diffblob) and merge (producing a mergeblob). Previously, there were many more steps of pure functions interspersed with database calls, but now those two steps just take a couple callbacks that load defnitions, names, whatever is needed.

I also deleted (& sort of inlined into Share) the "human diffs" calculation, as only Share uses it, so it doesn't need to be in the UCM side. I switched the "renames" field of the diff to only include "simple" renames (one name to one other name), since that's what we'll (try to) incorporate into merge.

A performance number: computing the diff for a small change to @unison/base, which includes an add and an update with 5 or so transitive dependents, used to take 13 seconds, and now takes ~0.2 seconds (and is instantaneous if repeated with the same inputs, because cache).

Test coverage

Existing transcripts, and manual testing.