[BUG] numba-cuda does not support pure CTK wheel users

[leofang]

#155 was only half-done for numba-cuda's wheel support story. During an offline meeting, we agreed to add optional dependencies so that users can do, e.g.

$ pip install numba-cuda[cu12]

to get a self-consistent CUDA 12 environment set up. This was not done.

This issue is compounded by the custom steps in setting up the wheel CI:

numba-cuda/ci/test_wheel_deps_wheels.sh

Lines 13 to 21 in 0587e40

	# TODO: Replace with rapids-dependency-file-generator
	python -m pip install \
	psutil \
	cffi \
	"cuda-python==${CUDA_VER_MAJOR_MINOR}.*" \
	"nvidia-cuda-runtime-cu12==${CUDA_VER_MAJOR_MINOR}.*" \
	"nvidia-curand-cu12==${CURAND_VER}.*" \
	"nvidia-cuda-nvcc-cu12==${CUDA_VER_MAJOR_MINOR}.*" \
	"nvidia-cuda-nvrtc-cu12==${CUDA_VER_MAJOR_MINOR}.*" \

which is not documented in the installation guide and does not mimic end user experience.

[leofang]

Looks like I can't assign to myself here 🙃

[bdice]

RAPIDS cuDF would also like to consume numba-cuda[cu12], to get CTK wheel dependencies that we manually specify today. We added those dependencies manually here: rapidsai/cudf#18686