Integrate CCCL runtime APIs into nvbench

[pciolkosz]

I noticed nvbench has nvbench::cuda_stream type, we should refactor it to use cuda::stream_ref instead. There could be more examples like that or just internals could use the new CCCL runtime APIs

[bernhardmgruber]

I fully agree, that this would be useful! After that, we could also simplify all of CCCL's benchmarks that take the nvbench stream and rewrap it in a cuda_stream passed to the environment.

[oleksandr-pavlyk]

cuda::stream_ref was introduced in CTK 12.2, CCCL 2.2. NVBench support matrix includes 12.0 and 12.1.

[pciolkosz]

Ahh, we might need to wait until CUDA 14.0 then