Speed up dependency solver preparation

[thomashoneyman]

While looking into #801, I started wondering if we've left easy solver performance improvements on the table. Indeed, we have! The previous PR made calculating the matrix cascade ~18x faster, and this PR makes solving an individual package considerably faster — anywhere from ~1.3x to ~80x faster. It doesn't require a lot of code; in fact, this is a small net reduction of the solver implementation, and the majority of the new code is improved tests and the benchmarking code.

In short, the solver previously found reachable versions by repeatedly merging dependency maps until they stopped changing. In this rework, we instead use a worklist, expanding each reachable package version once. It also waits until after pruning to initialize diagnostic data, and stops merging source information that would immediately be discarded.

Benchmarks

Solving the latest manifest's dependencies for every package in the most recent registry snapshot went from 82.92s to 37.62s, or 2.20× faster. All 1,176 complete plans were identical.

Seconds per batch of 1,176 solves	Before	After
Mean	82.92	37.62
Median	81.32	37.81
Observed range	78.87–90.19	36.74–38.14

The version index here is 5,573 manifests from the registry, and each package's dependencies are solved independently and without a compiler constraint. Individual workloads improve by different amounts. Halogen's dependency solve went from 72.56ms to 10.68ms; a synthetic batch with a propagation-heavy, 60-package chain only improved from 242.91ms to 227.98ms. (This at least shows we aren't regressing — there is no case I could put together in which the new implementation is slower).

Here are some benchmark results, visually:

The charts are showing means from minibench from 40 solves per process, and two processes per variant. The whiskers in the chart are the two means. As stated before, the resulting plans are the same.

The fixed snapshot uses registry revision d51cb8e and registry-index revision 23d1e62.

[thomashoneyman]

@f-f addressed your comments and updated the text of the PR accordingly; I didn't change the implementation besides bringing in the test helpers, so your original review still stands.

[f-f]

thank you!