Update Build Toolchain

[thalassemia]

This removes the deprecated distutils in favor of setuptools and replaces the setup.py CLI with pip and build.

Here is a table of environments where python stochastic_arrow/test/test_arrow.py ran without issue after installing either an sdist (platform-agnostic) or wheel (platform-specific) that I built on macOS 15.1.1, Python 3.12, Cython 3.0.11, and Numpy 2.1.3 with python -m build --sdist/wheel:

Operating System	Python version	Numpy version	Build type
Docker debian:latest	2.7	1.14.0	sdist
Docker debian:latest	2.7	1.16.0	sdist
macOS 15.1.1	3.12	1.26.4	sdist
macOS 15.1.1	3.12	1.26.4	wheel

It's pretty cool that this can still be installed on Python 2.7 with just a few tweaks. The last row suggests that building with Numpy 2.1.3 does not make the wheel incompatible with older versions of Numpy.

Also, I added a cibuildwheel GitHub Action that should automatically build and publish wheels for all common platforms and recent Python versions. It should run pytest on every wheel it builds so we won't accidentally upload a non-functional wheel.

I got the wheel working on the new free-threaded release of Python 3.13. Unfortunately, it was almost universally slower than the normal Python 3.13 both to build and to run pytest.

[thalassemia]

In the process of getting pytest to work with cibuildwheel, I realized that our current directory structure can lead to import issues when trying to test a wheel that you built and installed. Specifically, calling pytest inside the root directory will import the stochastic_arrow source files instead of the installed wheel.

To fix this, I decided to follow PyPA's recommendation to have all package source files inside a src directory. I could then move the test directory out to the top level and move the data directory into test. From there, I just had to modify a few tests to load data from the right place. Now, pytest works seamlessly regardless of whether it is using an editable install ("Pytest" GitHub Action) or an installed wheel ("Build and upload to PyPI" GitHub Action).

Also, I added a temporary job to release.yml to upload the built wheels to TestPyPI, a mock PyPI. You can view the results of that here and try installing using pip install -i https://test.pypi.org/simple/ stochastic-arrow.

[1fish2]

This looks great!

I fixed the CLI test runner although bugs remain in the --plot and --memory tests. See the commit message.

[thalassemia]

Thanks for the review and edits @1fish2! I noticed that the memory test in its current form fails on MacOS and changed it to assert memory_increases <= 10 instead of <= 1. Not sure why though, or if it's even worth digging into. The plot test is not included in our pytest test suite so I'm fine with that one failing.

[prismofeverything]

Huge improvement all around, thanks for bringing arrow into the modern age (!) This pattern is very useful - if it were worthwhile this could be the basis of a cython template that could be reused in the future (just had to go through a similar thing recently).

[prismofeverything]

I granted trusted publisher access to this repository for the package on pypi in the release.yml workflow, let me know if additional configuration is required there.