[FEA]: Optimize, Cleanup, and Expose CUB Thread-Level Reduction

[fbusato]

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.

1. Optimize. CUB Thread-Level Reduction is a pure sequential operation, leaving the opportunity for ILP optimizations with tree-reduction strategy. This optimization is not straightforward because some instructions provide a 3-way version depending on specific GPU archs
2. Cleanup. There are several overloading functions and the calls sequence is not easy to follow. In addition, the prefix version could inhibit vector optimizations.
3. Expose. Add Thread-Level Reduction in the cub:: namespace. Provide documentation and add related tests

Describe the solution you'd like

Optimize, Cleanup, and Expose CUB Thread-Level Reduction

Describe alternatives you've considered

No response

Additional context

No response