WIP: explain-mode for parts of arguments

[Zac-HD]

This currently-broken spike aims to extend #3555 (original issue: #3411) to report on sub-parts of arguments. Because this only makes sense for fixed-size collections - otherwise the minimal failing example would omit that part, so there'd be nothing to report on - we'll implement it for arguments to st.builds(), elements of st.tuples(), and for the non-optional keys inst.fixed_dictionaries().

Not a top priority right now, so I'm opening this just to mark where it's up to and ensure that we don't lose any progress.

---

Because this means we might have many more reportable spans to report on, it might also require some performance optimizations. The two obvious big ones are:

1. refactor ConjectureData to natively support vary-this-span, which would save up to 50% of runtime when we retry on changed span size
   • easier: for changed-length spans, we can skip the retry if it's shorter on the assumption that valid shrinks have already been tried. Maybe we can also use generate_novel_prefix() here, and exploit lexical ordering the same way?
2. reduce the number of failing examples we require before concluding that a part can vary.
   • this is currently 100, but we could accept a higher false-report rate to go faster
   • or incorporate a time-based heuristic so that we only make that trade when the runtime would otherwise be too long. For example, order the parts to check by nesting depth, abort at $timeout, and discard reports for the incomplete depth (or abort after finishing the current depth or a further timeout).

[Zac-HD]

Closing with intent to revive on top of #3864.