[CUB] Expose tuning parameters for `WarpReduceBatched`

Currently the implementation switches between WSPRO algorithm and REDUX instructions depending on `Batches` and `LogicalWarpThreads` according to our (artificial) benchmark heuristics. We could expose a minimum number of threads (currently 32) and maximum number of batches (currently 8) for which REDUX is used (if possible for the given type and reduction operator). Alternatively we could have a simple on/off/heuristic switch.