Numpy now uses a version cap for setuptools in their `pyproject.toml`,
which defies a bit the value of including it in the integration tests.

As revealed in a conversation with `numpy` maintainer, it seems that
they not plan to remove this cap (maybe only update it from time to
time, only if necessary). Moreover they are also studying other
build backends.

So I think that now the best thing to do would be stop trying to build
numpy sdists with the current version of setuptools, otherwise we risk
to break this test.

This is more reliable then simply globing
(and will also work in other operating systems)

Instead of disabling pytest plugins, simply don't install them

The selection of packages used in the integration test is arbitrary, and
can be changed.

The main criteria used was the time to build, and the number of
"non-Python" dependencies. The only exception was numpy, due to its
significance to the ecosystem.