[blocked] wheels: build with newest CUDA 13 CTK

[jameslamb]

Description

In RAPIDS 26.04, we switched to building CUDA 13 wheels with CUDA Toolkit 13.0: #257

This was done to avoid building against libnvJitLink 13.1, which would have brought in a runtime requirement of nvidia-nvjitlink>=13.1, which would have made RAPIDS wheels uninstallable against any available torch wheels (for CUDA 13).

Eventually we will want to switch back to building RAPIDS wheels against the latest available CTK. This tracks that work.

Benefits of this work

• take advantage of bugfixes, performance improvements, and new features in new CTKs

Acceptance Criteria

RAPIDS wheels:

• are built against the latest CTK at the time this is picked up
• all have appropriate runtime dependencies (e.g. nvidia-nvjitlink floor updated)
• every project with torch wheel tests uses CUDA torch wheels for CUDA 12 and 13 (ideally without needing pip install --no-deps to piece together the environment)

Approach

• ci-imgs changes
  • stop building CDUA 13.0.x ci-wheel images ()
• shared-workflows changes
• library changes, in RAPIDS dependency order
  • rmm
  • kvikio
  • dask-cuda not necessary: pure Python
  • ucxx
  • raft
  • cuvs
  • nvforest
  • cudf
  • cuopt
  • cucim
  • cuxfilter
  • rapidsmpf
  • cugraph
  • cuml
  • cugraph-gnn
  • nx-cugraph not necessary: pure Python
• ensure every project is using @main / @release/YY.MM of shared-workflows again

Notes

N/A