fix: don't try to re-wrap array_function overload results

[agoose77]

Our integration with NumPy's array function mechanism means that users can in some cases invoke NumPy-only functions on Awkward Arrays (if they're rectilinear). However, our existing support for this mechanism is breaking Awkward overloads such as ak.broadcast_arrays, because it assumes that the result should be a layout.

In the long run, we could apply an inverse _to_rectilinear transform to recover high-level ak.Array/ak.Records, but in this instance, I think the solution is even more clear cut.

If Awkward has an overload for an array function, then it should already be returning high-level objects. Therefore, we don't need to re-wrap the result. This PR applies such a fix, which means that any NumPy-only overloads that return non-arrays will still need a fix. However, this is a simple first step.

Fixes #2078

[codecov]

Codecov Report

Merging #2079 (135f2e6) into main (26f9481) will increase coverage by 0.00%.
The diff coverage is 100.00%.

Additional details and impacted files

Impacted Files	Coverage Δ
src/awkward/_connect/numpy.py	67.28% <100.00%> (+0.16%)	⬆️

[jpivarski]

Okay, got it. This is looking good so far. I'll take another look when it's out of draft mode (remind me).

Nope, it is out of draft mode! (I was thinking of a different PR.) I did take a good look at this and it makes sense. It is mergeable as-is, with the understanding that it's only for a simple case and that more is needed to solve the problem overall.

[agoose77]

Right. It doesn't try to solve the NumPy side of things, e.g. if we didn't define ak.broadcast_arrays, the result of calling np.broadcast_arrays would fail in the same manner.

We could be smarter there, but I also think there comes a point where we need to write an overload to handle the dynamicism of the NumPy API. That said, exploring returned tuples / lists and interpreting np.ndarray is definitely possible. Thankfully, we would still want this PR if we do decide to go that extra step in the future.