[Libxc_GPU] Upgrade build recipe for CUDA v13

[abussy]

Upgraded the build_tarball.jl script to accommodate CUDA v13.

From v13 on, NVIDIA does not ship NVVM as part of the NVCC redist. For a successful build on aarch64, NVVM must now be downloaded separately. The platforms/cuda.jl file was modified in order to make these changes available to other packages.

[ejmeitz]

I mean looks fine to me. Follows what I've done in the past to get things working on ARM. I do know @maleadt said using the redistributed sources was not the "right" way. That said I know what was done here works.

[maleadt]

All this code is bad and should go: https://github.com/abussy/Yggdrasil/blob/7426d9476148211fa13046f2670596fe00b9daa9/platforms/cuda.jl#L241-L326

Can you explain what the missing pieces are? For a downstream recipe like this, you should rely on any of our JLLs to provide the necessary build-time or run-time artifacts. Worst case, you can use the get_sources functionality to fetch things that aren't packaged by JLLs. But old-style manually maintained links to redist sources are not OK.

[abussy]

Originally, the code you highlighted (https://github.com/abussy/Yggdrasil/blob/7426d9476148211fa13046f2670596fe00b9daa9/platforms/cuda.jl#L241-L326) comes from the necessity of building CUDA code for the aarch64 architecture from BinaryBuilder (which runs on x86_64). This was introduced in #9910, and generalized from the Reactant recipe by @giordano . At that point, that was the accepted way for cross-compilation of CUDA code, and it has been adopted by other packages: Faiss, legate, MAGMA and a few others.

I am happy to try and get a working nvcc compiler through other means than an explicit redist download, but if I don't manage, I think this solution needs to stay.

[maleadt]

I am happy to try and get a working nvcc compiler through other means than an explicit redist download, but if I don't manage, I think this solution needs to stay.

In any case, you should use get_sources to automatically fetch stuff from the redist servers instead of hardcoding. A JLL-based solution is even better, of course, but may be tricky because of the platform mismatches.

[abussy]

I tried building with JLLs but did not manage, as the host architecture is x86_64-linux-musl, and the CUDA_SDK_jll are not distributed for musl. Therefore, using a HostBuildDependency is not possible.

Switched to get_sources from C/CUDA/common.jl to download the required redist. Other packages still used the cuda_nvcc_redist_source, but I think this refactoring should be its own PR.