CI Azure: try removing setuptools workaround on windows

[jeremiedbb]

Let's see if we still trigger the issue

[thomasjpfan]

I suspect this fails randomly on windows. For example the scheduled wheel builder job sometimes succeeds and sometimes fails on windows: https://github.com/scikit-learn/scikit-learn/actions/workflows/wheels.yml?query=branch%3Amain+event%3Aschedule

[jeremiedbb]

I think it's slightly different. The env var was to tell setuptools to use the legacy behavior of distutils because distutils introduced a breaking change without taking into account the packages that monkeypatch it like numpy. The issue we expected was not random: numpy/numpy#17216. Distutils released a fallback mechanism but in the mean time we needed this env var.

If I'm right, it does not mean that the windows build is fixed. It only means that the fallback mechanism works, but we might still be building with the not thread safe CCompiler, (i.e. we still need #22693)

[jeremiedbb]

The PR that introduced this is #22049 which seems to have been to quickly fix the CI because it was failing in every PR, which confirms it being not a random failure

[thomasjpfan]

I think there are two part of this issue:

1. setup.py running at all: The original error came from setuptools and numpy both monkeypatching disutils that were incompatible. This should be fixed since distuils added a faillback mechanism for the 'env' argument, which is vendored by setuptools.
2. setup.py compiling without failure on windows: From my understanding, we need a NumPy with this patch: BUG: Support env argument in CCompiler.spawn numpy/numpy#20650 to avoid rc.exe randomly not found (race condition?) pypa/setuptools#2212 .

Edit: This is exactly what you said 😅

[jeremiedbb]

Now I don't understand why we have deprecation warnings from distutils that we don't have in other jobs...
E DeprecationWarning: distutils Version classes are deprecated. Use packaging.version instead.

[ogrisel]

This deprecation warning is raised by:

C:\Miniconda\envs\testvenv\lib\site-packages\joblib\backports.py:36: in make_memmap

so this is a problem in joblib (not scikit-learn directly). I fixed it in a recently merged PR by backporting LooseVersion in joblib: https://github.com/joblib/joblib/pull/1272/files#diff-01a388465308c78a2759807da9e4aec087311a155205d483b957bb54382b0561 but this is not released yet.

[jeremiedbb]

But if it's not released yet, why don't we trigger that in the other CI jobs ?

[jeremiedbb]

@thomasjpfan I think we can merge this one