Draft: Port to uv + scikit core build

[rpep]

Depends on other PR

[rpep]

@davidcortesortuno This moves away from setup.py to more modern way of building Python packages - fully pyproject.toml-able, no Python code executed during build.

[rpep]

@davidcortesortuno still need to fix the Docker builds.

[davidcortesortuno]

I really like the idea of moving to uv, but the build process for the cython extensions looks more complicated than using setup.py now. Do you think it's better to use Cmake in terms of maintainability?

[rpep]

@davidcortesortuno it's just become more standard unfortunately/fortunately depending on your perspective. The plus side is that by using standard FindX stuff in CMake it can hook into pkgconfig and find the libraries using stnadard methods for Linux. It's more familiar for packagers. I was looking at trying to get Fidimag into Spack https://spack.io/ and with this sort of setup it's all quite easy as it's a well trodden route.

The main thing is that setup.py is now not seen as a good option; people don't want to run arbitrary Python code at pip install time. CMake is also significantly better at handling different compilers and toolchians than our hacked-together setup was.

If you try this (obviously you need CMake installed) you see that it'll recompile automatically on change too, it's quite seamless.

[davidcortesortuno]

Fair enough. I was also thinking if we should move away from Cython as well, in favour of something like nanobind, since most of the Cython code is just wrappers to call pure C functions, with loads of repeated function headers (instead of using pure Cython)

I recently saw that Sundials 7.6 started to provide experimental official Python bindings https://sundials.readthedocs.io/en/latest/Python/Usage.html This would allow us to drop our sundials pyx file. Since you already updated Sundials to the latest version I recommend we should try it. I guess this will ease the compilation part of the modules

[rpep]

PyBind11 is very popular but I've never tried it personally. I'm not sure the Cython stuff matters given the overhead of calls will be tiny compared to the C function calls.

[rpep]

Re: Sundials Python bindings - maybe as another stage after this? Don't want the PR to grow much more :D

[davidcortesortuno]

Re: Sundials Python bindings - maybe as another stage after this? Don't want the PR to grow much more :D

Sure!