[CUB] Make warpspeed stable reduction default implemenation and discard unstable version

https://github.com/NVIDIA/cccl/pull/9263 PR makes warpspeed scan run_to_run deterministic, by having a stable reduction tree always. Because of which we have two implementations which with very minor change.

Can we just keep the stable reduction tree version as default and discard unstable version ?

If yes, We have to address two key problems before doing that
- [ ] Confirm there is no performance regressions on all GPUs and data-types we support warpspeed today
- [ ] Do we have to separate the implementations if for some GPUs in future there is performance divergence ?