Update to CCCL 2.8

[bdice]

This issue tracks the progress of updating RAPIDS to CCCL 2.8.0. The tag v2.8.0 is now available: https://github.com/NVIDIA/cccl/releases/tag/v2.8.0

[bdice]

We may want v2.8.2. Version v2.8.1 was tested and showed a regression on GCC 13 with Wmaybe-uninitialized. See NVIDIA/cccl#4054 for details. That PR is now merged, so I will start a CCCL migration to 2.8.x and pin to v2.8.2-rc0.

[bdice]

Status:

• RMM: passing ✅ (Update to CCCL 2.8.2-rc0. rmm#1855)
• KvikIO: passing ✅ (Update to CCCL 2.8.2-rc0. kvikio#658)
• cuDF: some test failures ❌, rerunning after merging branch-25.04 (Update to CCCL 2.8.2-rc0. cudf#18235)
  • Appear to be off-by-one errors
  • Need to check on the CCCL patches too!
• RAFT: passing ✅ (Update to CCCL 2.8.2-rc0. raft#2603)
• cuVS: some test failures ❓, rerunning because the tests are probably flaky (Update to CCCL 2.8.2-rc0. cuvs#757)
• cumlprims_mg: passing ✅ (https://github.com/rapidsai/cumlprims_mg/pull/235)
• cuSpatial: some test failures that are unrelated, rerunning❓(Update to CCCL 2.8.2-rc0. cuspatial#1548)
• cuML: some test failures ❌ (Update to CCCL 2.8.2-rc0. cuml#6421)
  • Appear to be off-by-one errors
• cuGraph: some test failures, rerunning after merging branch-25.04 ❌ (Update to CCCL 2.8.2-rc0. cugraph#4967)
  • C++ exception with description "transform: failed inside CUB: cudaErrorInvalidDevice: invalid device ordinal" thrown in the test body.

[miscco]

We definitely need to check whether the patches still apply

C++ exception with description "transform: failed inside CUB: cudaErrorInvalidDevice: invalid device ordinal" thrown in the test body.

Looks like an infra issue