feat: implement sum, count, and count_nonzero reductions

[agoose77]

@jpivarski @Saransh-cpp this is a WIP initial stab at supporting sum, count, and count_nonzero reducers in vector across both NumPy and Awkward.¹

My main query here is around API design; I note that @jpivarski and @henryiii designed the initial prototype, so you'll both likely have stronger feelings on how this deviates from the intended design. In particular, is it reasonable to fit reducers (in this case, only sum) into the compute model that is already established? From our type hints here, we don't explicitly encode any expectation that the vectors in compute functions are array like (or indeed satisfy any type checks), so I think this is fine.

Awkward support depends upon scikit-hep/awkward#2458.

Footnotes

1. count is Awkward specific, however. ↩

[codecov-commenter]

Codecov Report

Patch coverage: 68.42% and project coverage change: -0.10 ⚠️

Comparison is base (8ca75f6) 83.01% compared to head (f964f97) 82.91%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #347      +/-   ##
==========================================
- Coverage   83.01%   82.91%   -0.10%     
==========================================
  Files          96       96              
  Lines       11372    11435      +63     
==========================================
+ Hits         9440     9481      +41     
- Misses       1932     1954      +22     

Impacted Files	Coverage Δ
src/vector/backends/awkward.py	81.21% <51.35%> (-1.83%)	⬇️
src/vector/backends/numpy.py	81.50% <84.61%> (+0.37%)	⬆️

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[agoose77]

Here's a rewrite that implements the reductions for each array backend. I need to cleanup the type hints, but it's mostly there.

[Saransh-cpp]

This should also close #92 and maybe iris-hep/func_adl_uproot#98 🎉

PS: I had a very dirty somewhat working implementation for np.sum here - #92 (comment)

[agoose77]

This should also close #92 and maybe iris-hep/func_adl_uproot#98 🎉

PS: I had a very dirty somewhat working implementation for np.sum here - #92 (comment)

Oops, I missed this! Well, good to know it's needed!

[agoose77]

NB, this PR took a journey following feedback from Jim. In its current (final!) form, we don't worry with coordinate-systems; everything is moved into Cartesian, and remains in Cartesian.

[jpivarski]

everything is moved into Cartesian, and remains in Catesian

Some (many?) of the ufunc-like functions do, too. This is not unique to reducers.

[agoose77]

This is ready to go!

[jpivarski]

This is ready to go now because Awkward 2.2.3 is out. (The tests would have failed if they couldn't reach it.)

I did another scan through the diffs, and I'm reminded that this includes some bug-fixes as well as the new feature, but I guess that's alright. (There's assignment of __numba_typer__ and __numba_lower__ into Vector's behavior instead of ak.behavior, and also type hint corrections on allclose—booleans, not floats—that I hadn't noticed before.)

I'll merge it.