SciPy complained about "expression result unused" for the extra '(',
which makes sense.

It also complained about `rc` being unused. This is slightly annoying
because we put the return value of `prima_minimize` in `result.status`.
I decided it made sense to return info by itself if there was an error
in either `init_result` or `check_problem`, and then just return DFT by
default.

- gcc 13.3.0 seems to have an issue with math.h on macOS. It spews a
  bunch of errors about the API_DEPRECATED macro not being used
  correctly, which is not an issue on our side. gcc 13.2.0 does not
  seem to have this issue, but since we cannot select gcc with such
  granularity we downgrade to gcc 12 and hope this fixes the issue for
  now.

- numpy 2.0.0 has come out which causes runtime issues. Fortunately the
  fix is easy, just build with pybind11 >=2.12.0. However this is still
  an issue for the pdfo compatibility test since pdfo does not support
  numpy 2.0.0, and sometimes the tests install 2.0.0 because it happens
  to be in the pip cache. So we upgrade pybind11 (the new version is
  compatible with 2.0 and 1.0 numpy) and skip the pdfo test if we
  happened to pick up numpy 2.

- macOS builds for building Python wheels started to complain when running
  the "delocate" step that libquadmath, libstdc++, and libgfortran have
  minimum targets of 11.0, whereas the default was 10.9, hence we added
  the relevant env variable for CIBW.

- Notes have been made in build_python.yml for failures related to
  Ubuntu 24.04.

- A couple tests were failing on macOS 14. I tried to modify rhobeg and
  rhoend, but this ended up turning into a game of whack-a-mole, where a
  rhobeg/rhoend that worked on one architecture would fail on another.
  Ultimaately the most reliable solution I found was to modify the
  starting point to be close to the optimal point.