Make Flaky into an ExceptionGroup

[jobh]

This PR makes Flaky into an ExceptionGroup to improve presentation of flaky failures. Closes #4040.

As usual, There Were Complications. Basically because an ExceptionGroup is not allowed to have an empty group; so a straightforward replacement is not possible.

On the plus side, I think the reworked Flaky is more expressive by differentiating two different kinds of flakiness (replay inconsistencies vs. inconsistent data generation), and it does improve reporting of flaky errors detected by the engine by adding full exceptions on the way out.

HypothesisException (via _Trimmable)
    |
    +Flaky              Base class for all flakiness (not currently raised anywhere).
       |
       + FlakyData      Raised to user. Conjecture detects inconsistent data generation.
       |
       + FlakyReplay    Internal. Conjecture detects flaky replay failure, but has only InterestingOrigins.
       |
       + FlakyFailure   Raised to user. ExceptionGroup, possibly constructed from a FlakyReplay.

A simpler flat alternative is this, with the only drawback being that InconsistentData (renamed from FlakyData) is no longer Flaky. I think I actually prefer this version, but I don't know how you feel about this drawback — is it even acceptable for a minor revision, do users ever catch Flaky? — so the implemented version is the one above.

HypothesisException (via _Trimmable)
    |
    + Flaky                Raised to user. ExceptionGroup, possibly constructed from a FlakyReplay.
    |
    + FlakyReplay          Internal. Conjecture detects flaky replay failure, but has only InterestingOrigins.
    |
    + InconsistentData     Raised to user. Conjecture detects inconsistent data generation.

[tybug]

is it even acceptable for a minor revision, do users ever catch Flaky?

Looks pretty rare, with the majority being schemathesis: https://github.com/search?q=%22except+Flaky%3A%22+OR+%22except+Flaky+as%22+language%3Apython&type=code

That said, I definitely have some local research code that catches Flaky...but I knew I was flying close to the sun with that one.

[Zac-HD]

Seems rare, but while I share your preference for the flat version I think we should continue to support Flaky - power-users and researchers are still users, and I'm generally comfortable taking on some more maintainece burden to make using Hypothesis delightful 🙂

[jobh]

So be it 😇 Ready for review then! No rush - I'm starting my summer vacation today, and will be on and off depending on weather. Hopefully mostly off 😁

[Zac-HD]

I want to play around and compare the before-and-after reporting for some examples before we merge, but this looks great so far 😁

[Zac-HD]

I'm concerned that users will think we're reporting an internal error rather than something they need to fix, but I can't think of any name that wouldn't 😕

[jobh]

Naming is hard! I went through FlakyGeneration plus some variants of InconsistentX and IndetermisticX, and not at all sure I settled for the best.

[jobh]

FlakyStrategy? And rename the other one to FlakyExecution, for symmetry?

Edit: maybe it's bad form to like my own comment, but I think these names would better. FlakyExecution and FlakyStrategy.

[Zac-HD]

I do like FlakyReplay as a concise and evocative name; for the other I've gone with FlakyStrategyDefinition which is more verbose but does seem to point to the thing-which-is-flaky.

[Zac-HD]

Played around a bit more and I'm very happy with this; I've also pulled out a cross-backend-specific thing from #4034 which will help us and our users debug such issues.

Thanks again, @jobh!

[Zac-HD]

I do like FlakyReplay as a concise and evocative name; for the other I've gone with FlakyStrategyDefinition which is more verbose but does seem to point to the thing-which-is-flaky.