CUDA support

[K20shores]

We have some hacked-in cuda support that works for linux. I've found a package that seems to also support it for windows

https://github.com/isl-org/Open3D

We should consider if we should adopt their approach, and maybe also see if they have a better way of building the python library with cuda support. Seems like they have two pacakges, one with and one without gpu support. It is quite a bit of work for us to optionally have cpu/gpu in one package build, depending on platform and availability of nvidia libraries. Something to consider

[K20shores]

#645 renamed musica to python to separate the bindings. With this, I realized that our cpp files seem to packaged alongside the python code. There's really no reason to do this. We could setup the build similar to Open3D by moving the c++ files for our python to src/pybind or something

@boulderdaze what do you think about organizing the files that way?

[boulderdaze]

Yeah I'm fine with that approach

[K20shores]

Some updates:

• Correct python packaging #735 needs to happen because wheels like ours will be rejected starting february 1st, 2026
  • We are rejected because our record file for linux does not match the contents. I think this is because of how we have to repair wheels for cuda support
  • If we figured out Refactor MUSICA to Support Optional Runtime CUDA via Type-Erasure #723, we would probably have an easier time...
• There may soon be support for wheel variants which would allow us to build the wheel multiple times for features like cuda support, making our weird double build no longer necessary
  • scikit-build-core may include support for this

[K20shores]

and because I always want this list but can never remember where it is

• cibuildwheel list of projects and platforms they support
• relevant PR with examples of packaing cuda testing python cuda build #305