New issue by tornaria on void-packages repository

https://github.com/void-linux/void-packages/issues/36062

Description:
For some (unspecified) reason the numpy template contains this in its `post_install()`:
```
	# create compat symlinks for .h files
	vmkdir ${py3_inc}
	ln -sfr ${DESTDIR}/${py3_sitelib}/numpy/core/include/numpy \
		${DESTDIR}/${py3_inc}
```

This creates a symlink `/usr/include/python3.10/numpy -> ../../lib/python3.10/site-packages/numpy/core/include/numpy`

In effect this means that whenever one uses `-I/usr/include/python3.10` to add python headers, one is also adding system numpy headers.

The problem arises when one uses a different version of numpy installed in a venv. To compile against numpy headers one would add `-I$VENV/lib/python3.10/site-packages/numpy/core/include` to use those headers. However, when the `CFLAGS` has something like `-I/usr/include/python3.10 -I$VENV/lib/python3.10/site-packages/numpy/core/include` the symlink installed in the python include dir will take precedence.

This gives a runtime error, e.g. when system numpy is 1.22 (as in void) and venv numpy is 1.21 (as in sagemath) since those versions are binary-incompatible (the numpy.ndarray structure changed size).

A workaround is to ensure the numpy include dir always comes before the python include dir, but the issue is bound to resurface since controlling the order of stuff placed on CFLAGS is kind of difficult.

As a concrete example: sage-9.5 (current release) used to build ok on void linux but after numpy upgraded to 1.22 it no longer builds (this does NOT affect our template for sagemath -- we don't use a venv). See: https://trac.sagemath.org/ticket/33473.

Arguably, nothing else than python headers should be installed in `/usr/include/python3.10`; installing headers from other packages  there is forcing the use of a particular version of the package.

Note: debian/ubuntu also adds this symlink. I didn't find other distros that do  (I only looked at a few).