[ENH]: Improve/Rewrite `cuda.parallel`'s build system

[leofang]

Building cuda.parallel is quite brittle due to requirements from the C library. Through some patient trials and errors I discovered the following build-time dependencies are required:

• gcc 13+ (due to the usage of std::format)
• CUDA 12.5+ (due to the usage of gcc 13, which is otherwise unresolvable with conda-forge CUDA packages)
• make (this is a quirk of conda-forge)
• libnvjitlink-dev (needed to pass CMake checks)
• cuda-nvrtc-dev (needed to pass CMake checks)
• lit (needed to pass CMake checks)

Also I have to set the env var CUDAARCHS like CUDAARCHS="86;89" pip install -v . so that CMake knows which archs to build for, otherwise CMake also complains.

Suggestions:

• Migrate the build system from setuptools + custom build commands to more CMake-friendly scikit-build-core
• Add build-time requirements to the section in pyproject.toml for scikit-build-core to pick up
• Skip unnecessary CMake checks if only building the C library
• Figure out a way to build also for CUDA 11 (for which the underlying C library is supposed to work) so as to support deployments based on CUDA minor version compatibility
• Set a default CUDA arch value and allow users to overwrite
• Test conda builds in the CI like the RAPIDS projects do
• Expand the installation section to add instructions for building from source

[leofang]

I noted that due to the way we look up headers at run time, editable install (pip install -e .) does not work.

[miscco]

I have opened a PR to address:

• [ ] gcc 13+ (due to the usage of std::format)
• [ ] CUDA 12.5+ (due to the usage of gcc 13, which is otherwise unresolvable with conda-forge CUDA packages)

I believe we can get rid of

• [ ] make (this is a quirk of conda-forge)
• [ ] lit (needed to pass CMake checks)

AFAIK we need

• [ ] libnvjitlink-dev (needed to pass CMake checks)
• [ ] cuda-nvrtc-dev (needed to pass CMake checks)

[gevtushenko]

Issue tracking CCCL headers installation in cuda.parallel: https://github.com/NVIDIA/cccl/issues/2281 Issue tracking CMake improvements for CCCL/c: https://github.com/NVIDIA/cccl/issues/2235