[FEA]: Combine Three-Way Partition Look-backs

[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.

At the time of three-way partition implementation, it was faster to perform two look-backs of I32 counter to figure out the number of items selected by op_1, and op_2 compared to performing one I64 look-back of a counter pair. Since the look-back was recently improved, we should revisit this decision.

Describe the solution you'd like

We should try combining two look-backs into a single one, tune it and compare resulting performance of three-way partition.

Describe alternatives you've considered

No response

Additional context

No response