Add nvidia-nvjitlink to pip dependencies for packages that depend on nvidia-cusparse

[bdice]

Problem

The PyPI nvidia-cusparse-cu12 package depends on nvidia-nvjitlink-cu12, but the dependency is not version-pinned. This means pip can resolve an older nvidia-nvjitlink-cu12 than what nvidia-cusparse-cu12 actually requires at runtime, leading to errors like:

undefined symbol: _nvJitLinkGetErrorLogSize_12_9

Root cause

RAPIDS libraries declare cuda-toolkit[...,cusparse,...] in their pip dependencies via dependencies.yaml, but do not explicitly include the nvjitlink extra. Since nvidia-cusparse's dependency on nvidia-nvjitlink is unpinned on PyPI, pip's resolver may select an incompatible (older) version of nvidia-nvjitlink.

Fix

Add nvjitlink to the cuda-toolkit[...] extras in dependencies.yaml for all affected repos, ensuring pip installs a version-compatible nvidia-nvjitlink wheel alongside nvidia-cusparse.

Affected repos

• rapidsai/raft
• rapidsai/cuvs
• rapidsai/cuml
• nvidia/cuopt

[bdice]

PRs:

• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility raft#2948
• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility cuvs#1794
• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility cuml#7794
• Add nvjitlink to cuda-toolkit pip extras for cusparse compatibility NVIDIA/cuopt#850