refactor: add array comparison test helper

[agoose77]

Add a helper function to ak._util that makes approximate comparisons of ragged arrays.

Some of the branches in this helper may not have been tested; my view is that we'll discover any bugs when we use it in a test. I've checked all of the primitive cases, though.

[codecov]

Codecov Report

Merging #2024 (7ede827) into main (b83d9dc) will decrease coverage by 0.00%.
The diff coverage is 74.35%.

Additional details and impacted files

Impacted Files	Coverage Δ
src/awkward/_util.py	80.79% <74.35%> (-0.51%)	⬇️

[jpivarski]

Note for posterity: we talked about this on Slack. But there are a few (good) additions since then, too.

You extract the two behaviors, but do you check them for equality? If so, I don't see it.

The "approximate dtype" is interesting: normally, we want types to be exact, but we do have an issue in which NumPy on Windows initializes integers as 32-bit. In our tests so far, we try to be diligent about explicitly setting the integer dtype to int64 when using NumPy, but this function would accept 32-bit integers and 64-bit integers as the same.

Overall, I approve! We'll probably be tweaking it as we use it, to more clearly define what we want to call equal same what we want to call not-equal.

[agoose77]

You extract the two behaviors, but do you check them for equality? If so, I don't see it.

No, I only check that they map to the same behaviour class for the array / record parameters. This might not be sufficient; we obviously also permit overloads of ufuncs, which wouldn't be caught this way. Equally, I wouldn't want to ban cases where additional behaviour classes were added to the behaviour dictionary that aren't used. Unless we consider that an desirable outcome?

The "approximate dtype" is interesting: normally, we want types to be exact, but we do have an issue in which NumPy on Windows initializes integers as 32-bit. In our tests so far, we try to be diligent about explicitly setting the integer dtype to int64 when using NumPy, but this function would accept 32-bit integers and 64-bit integers as the same.

Right, I'm worried about

arrays_approx_equal(array, [1, 2, ...]) 

failing due to the default types for ArrayBuilder, Perhaps that should be opt-in, though, not opt-out.