[FEA]: Support device-resident problem size in device reduce

[gevtushenko]

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.

Currently, cub::DeviceReduce accepts num_items as an integer. Sometimes, problem size is not available at host side.

Describe the solution you'd like

This issue can be closed with a PR adding device-resident problem size support in device reduce:

int *num_items = ...;
cub::DeviceReduce(...,  cuda::args::deferred{num_items})

Test should include passing a fancy (say, transform) iterator into deferred.

Describe alternatives you've considered

No response

Additional context

No response

[bernhardmgruber]

Seems very related to: #9444

[gevtushenko]

Seems very related to: #9444

#9444 is related by focused on initial value. This one is for problem size.