Parallelize definition, reference, and reference result passes

[efritz]

This makes the traversal of packages for the definition, reference, and reference result passes happen concurrently (up to the number of logical CPUs on the index machine).

The logic changes are mostly in visit.go, which now surrounds the loops with a goroutine that instead feeds a channel for runParallel. The remaining work (the majority of changes) is surrounding access of shared datastructures with appropriate mutexes. This is just a sync.[RW]Mutex in most cases, except for the multi-level ranges map, which is guarded by a striped mutex.

This implementation is similar to https://github.com/nmvalera/striped-mutex, but does not require hashing the string key to convert it into a stripe index.

Each commit is a logical group, so reviewers may find it easier to review commit-by-commit rather than look at the resulting diff.

Improvement:

[aidaeology]

LGTM; very excited about the speed improvements here 🥇