fix detection of math library in numpy build

[Flamefire]

(created using eb --new-pr)

NumPy detects the required math by trying to link a minimal code containing a call to log(0.).
The first try is without any libraries, which works with gcc -fno-math-errno (our optimization default)
because the call gets removed due to not having any effect. So it concludes that -lm is not required.
This then fails to detect availability of functions such as acosh which do not get removed in the same
way and so less exact replacements are used instead which e.g. fail the tests on PPC.

$MATHLIB=m makes it try -lm first and is supported until the Meson backend is used in 1.26+.

The test code used is:

double log(double);

int main(){
  log(0.);
}

The issue can be reproduced by putting it in a file test.c and running:

$ gcc test.c
# OK
$ gcc -fno-math-errno test.c
/software/.../bin/ld: /tmp/ccGF6GFx.o: in function `main':
test.c:(.text+0x20): undefined reference to `log'
collect2: error: ld returned 1 exit status

On PPC I see test failures in e.g. SciPy-bundle-2023.02-gfbf-2022b.eb & SciPy-bundle-2023.07-gfbf-2023a.eb
FAILED core/tests/test_umath.py::TestComplexFunctions::test_loss_of_precision[complex256] - AssertionError: (120, 1.78095631306534887655740488957533e-10, 5.2863423150802063925011506713427e-21, 'arcsinh')

However the issue is not limited to PPC although the fallback functions seem to be "good enough" on x86.
A minimal example I made: https://godbolt.org/z/zGYcW45qE and one that it fails to detect casinh without -lm