[FEA] Update to CCCL 2.3 or 2.4

[bdice]

Currently RAPIDS is built with CCCL 2.2. This issue lists tasks that we can follow up on once we have upgraded to CCCL 2.3 or 2.4. (We haven't decided on an exact timeline for updating, so RAPIDS could target CCCL 2.3 or 2.4 depending on that timing.)

CCCL 2.3

• If we upgrade to 2.3, we will need a patch for Fix issues with ambiguous calls to addressof in thrust::optional NVIDIA/cccl#1499 which will be fixed in 2.4.
• CCCL 2.3 performance may be a motivating factor:
  • Up to 60% performance improvements of cub::DeviceSelect::UniqueByKey, cub::DeviceScan::ExclusiveSumByKey, and cub::DeviceReduce::ReduceByKey on A100. cub::DeviceSegmentedReduce now supports 64-bit indexing.
• Replace device uses of thrust::optional with cuda::std::optional
  • Replace usages of thrust::optional with std::optional #15091 (comment)

CCCL 2.4

• See notes on patch above

Additional Context

• https://github.com/NVIDIA/cccl/releases
• Test PR: Test CCCL 2.3.0. #14704

cc: @miscco @jrhemstad @robertmaynard

[robertmaynard]

CUDA 12.4 ships with CCCL 2.3.

The CCCL support matrix states that CCCL 2.2 is officially untested/unsupported when using 12.4+, so moving to a newer version will be needed.

[robertmaynard]

CCCL 2.3 also requires changes to projects due to thrust::tuple constructor changes.

[bdice]

We can also tackle NVIDIA/cuCollections#469 after this migration is complete.

[vyasr]

Are we now considering going straight to 12.5? See rapidsai/rapids-cmake#607 (CC @trxcllnt).

[bdice]

RAPIDS has been updated to CCCL 2.5 (a pre-release commit for now, and eventually the 2.5.0 release).