[FEA]: Add runtime-cooperative segmented-reduce kernel and kernel picker

[anikaj-eng]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

CUB

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

The runtime-cooperative segmented-reduce experiment selects a cooperating group width at launch time, but its kernel, launcher, temporary-storage handling, and picker are not CUB building blocks.

Describe the solution you'd like

• Add the runtime-cooperative segmented-reduce kernel and launcher for the supported threads-per-segment widths.
• Add the picker that chooses the cooperative path or the CUB fallback from launch-time problem properties.

Close once these internal building blocks are implemented and tested.

Describe alternatives you've considered

No response

Additional context

Experiment references:

• benchmark flow
• launcher
• temporary-storage query
• picker

[anikaj-eng]

cc: @nvjonwong

[nvjonwong]

I'll take a look at it.