[ENHANCEMENT]: Make cuco code NVRTC-compatible

[PointKernel]

Is your feature request related to a problem? Please describe.

Related to NVIDIA/cccl#4305

Make sure cuco is JIT or NVRTC compatible.

Describe the solution you'd like

To enable cuco code to be used in JIT kernels, we need to ensure that no thrust algorithms are used in cuco implementations. When available, we will replace the two existing uses of thrust algorithms with their cuda::std:: equivalents.

cuCollections/include/cuco/detail/open_addressing/open_addressing_ref_impl.cuh

Line 1269 in e3aec27

thrust::reduce(thrust::seq, num_matches, num_matches + bucket_size);

cuCollections/include/cuco/detail/open_addressing/open_addressing_ref_impl.cuh

Line 1257 in e3aec27

if (thrust::any_of(thrust::seq, exists, exists + bucket_size, cuda::std::identity{})) {

As a simpler initial task, we can start by replacing thrust utilities such as pair, tuple, and equal_to with their corresponding cuda::std:: equivalents, which are already available for use.

[sleeepyjack]

How does rapids test for nvrtc compatibility? Can we set up a CI job for this?

[lamarrr]

Rapids CUDF has a couple of NVRTC-compatible kernels, we run these kernels as part of our tests via JITIFY.
Yes, you can set up a CI job for it; you'd need to use JITIFY2 or NVRTC directly instead of NVCC.
As long as the tests compile and run successfully, you'd be good.

[gengar-maker]

Hi! This ticket is still open. Is there any known problems with using NVRTC for cuco code? Or it's already working fine?

[lamarrr]

I haven't tested yet. Perhaps @PointKernel has?