Investigate regression of warpspeed scan on sm120

Copy & paste from a Slack message by @gevtushenko:

> I've benchmarked some of the scan algorithms on A6000 Pro, already see 40% ! speedup for 2^28 elements on:
>
> * cub.bench.scan.applications.P1.log-cdf-from-log-pdfs.base
> * cub.bench.scan.applications.P1.scan-over-unitriangular-group.base
> * cub.bench.scan.applications.P1.non-commutative-bicyclic-monoid.base
>
> That said, I see some regressions on 2^28 elements for:
>
> * 120% ! regression on cub.bench.scan.exclusive.custom.base on I8 (both I32 and I64 offsets)
> * ~10% regression for all types on cub.bench.scan.applications.P1.running-min-max.base
>
> The 10% regressions is not that concerning, giving that we provide far better speedups for other cases. But 120% seems suspicious.

Those regressions were introduced in #6811, which is heading for CCCL 3.4. We should investigate the regressions and solve them before 3.4 ships.