[FEA] Use CCCL random engines

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

https://nvidia.github.io/cccl/unstable/libcudacxx/standard_api/numerics_library/random.html#cuda-std-random

CCCL has a newly implemented STL-compliant, device-friendly random engine. We could refactor our benchmark data generators to use this and drop our dependency on libcurand.

Benefits:
- Might be easier to write data generators without custom kernels to call curand device functions?
- Fewer library dependencies is better