Argmin/argmax CUB comparators reject mixed tuple/pair inputs

[Aminsed]

MatX's optional build inside CCCL is currently failing when compiling the argmin/argmax reductions (see NVIDIA/cccl#6208 (comment)). CUB now instantiates the reduction functor with an accumulator stored in a while the input iterator provides a . / still take a single template parameter, so is ill-formed and the build stops in .

Repro:

1. Build MatX against current CCCL main (or run CCCL's optional MatX workflow).
2. Compilation fails in with the error above.

Expected: the comparators should accept heterogeneous argument types (pair vs tuple) and always return the tuple type that CUB expects.

This issue tracks adding the necessary overloads/helpers so MatX builds against latest CCCL.