[libcu++] Investigate minor performance regressions with `cuda::std::sort`

We found some minor performance regressions compared to thrust. Those are probably related to using larger difference type but we need to investigate

Please create a tracking issue to investigate the small performance difference to Thrust.

_Originally posted by @bernhardmgruber in https://github.com/NVIDIA/cccl/issues/8621#issuecomment-4428988450_