ci: fix oldest-supported-dependencies test workflow

[agoose77]

It turns out that our CI wasn't properly testing the package combinations that we'd asked it to — there were subsequent install steps that overwrote the intended variations. This means that #2767 was not properly tested against older versions of NumPy, leading to a bug.

Although it's possible to use multiple install steps, it best reflects our intentions if we use a single install command.

In future, we probably just want to have a single requirements.txt file, and a separate constraints.txt that pins these dependency versions on our lower-bound configuration. However, we don't currently test all dependencies at the lower bound. This PR only runs core tests (NumPy + PyArrow) in the 3.8 tests.

Fixes #2784

[codecov]

Codecov Report

Merging #2783 (14aebfd) into main (3f623ee) will decrease coverage by 0.08%.
The diff coverage is 57.53%.

Additional details and impacted files

Files	Coverage Δ
src/awkward/_nplikes/numpylike.py	73.72% <ø> (ø)
src/awkward/_nplikes/cupy.py	36.78% <33.33%> (-0.43%)	⬇️
src/awkward/_nplikes/array_module.py	84.15% <66.66%> (-3.35%)	⬇️
src/awkward/_nplikes/typetracer.py	75.33% <50.00%> (-1.26%)	⬇️

[agoose77]

414f69a should break the CI, such that 6c708c6 fixes it!

[jpivarski]

This PR fixes both the test infrastructure and also the bug for old versions of NumPy. I strongly approve of the use of a NUMPY_HAS_NEP_50 instead of finding a solution that works for both old and new NumPy, so that we know what to remove when our minimum NumPy does get to be 1.24 or above.

I think this is ready to merge!