Switch RAPIDS to static linking of libcudart by default

[bdice]

Summary

This issue tracks the effort to switch all RAPIDS libraries from dynamic linking (cuda-cudart) to static linking (cuda-cudart_static) of the CUDA runtime by default. This is an implementation of #4.

Motivation

• Simplifies deployment by embedding CUDA runtime into binaries
• Eliminates runtime dependency on cuda-cudart conda packages
• Reduces potential for CUDA version mismatches at runtime

Changes Required Per Repository

1. CMake: Set CUDA_STATIC_RUNTIME=ON as default in cpp/CMakeLists.txt
2. Build scripts: Update build.sh default to CUDA_STATIC_RUNTIME=ON
3. Conda recipes: Remove cuda-cudart from run requirements (keep cuda-cudart-dev in host)

Rollout Order (Bottom-Up by Dependency)

Phase	Repositories
1	rmm
2	kvikio, raft, cuvs, cudf
3	cumlprims_mg, cuml, cugraph, ucxx, rapidsmpf
4	cugraph-gnn, cucim, cuopt

Tracking

• rmm - Default to static linking of libcudart rmm#2178
  • Reverted due to breaking cuDF Python stream tests, @vyasr is investigating
• kvikio - Default to static linking of libcudart kvikio#885
• raft - Default to static linking of libcudart raft#2890
• cuvs - Default to static linking of libcudart cuvs#1627
• cudf - Default to static linking of libcudart cudf#20814
  • Blocked by cudf stream testing
• cumlprims_mg - Default to static linking of libcudart cumlprims_mg#30
• cuml - Default to static linking of libcudart cuml#7580
• cugraph - Default to static linking of libcudart cugraph#5366
• ucxx - Use static linkage for CUDA runtime ucxx#559
• rapidsmpf - Use static linkage for CUDA runtime rapidsmpf#725
• cugraph-gnn - Use static linkage for CUDA runtime cugraph-gnn#369
• cucim - Use static linkage for CUDA runtime cucim#992
• cuopt - Use static linkage for CUDA runtime NVIDIA/cuopt#680
• rapids-cmake - Default to static linking of libcudart rapids-cmake#949