Improve installation tests.

[1uc]

IIUC during the release of 8.2.4 we found defects, after passing CI. The problem seemed to be that during CI we always install setuptools because we need it to build the package. Therefore, when we test that the package if fully functional setuptools happens to already be installed; and it works even if setuptools is not part of the runtime-requirements and wouldn't be installed by installing neuron. Since, we need setuptools at runtime and Python 3.12 stopped shipping it, we can't rely on it being installed automatically on users machines.

We should review the CI that test functionality of our wheels. Maybe it needs to split building and installing + testing more rigorously.

@JCGoran, @heerener please add any relevant information and corrections.

[heerener]

Note: this comment was not driven by deep research, only a quick glance and my own experience building/testing wheels.

When I build and test wheels on my macbook, I'm generally left with a bunch of virtualenvs: one per python version for building, and an additional one per python version for testing.

https://github.com/neuronsimulator/nrn/blob/master/packaging/python/test_wheels.sh#L245 shows that the test_wheels.sh script creates a separate virtualenv for testing.

It could still be the case that something in the pipeline still reuses a venv somewhere, but there may be something else going on.