Hierarchical Parallelism improvements

[mrnorman]

Create inner level parallel reductions using CUB and hipCUB inside the YAKL_reductions.h header, and create simple intrinsics functions for sum, maxval, minval, and product.

Whenever inner-level loop size exceeds the launch bounds size, strip-mine the loop. This allows more flexibility and fewer concerns for the user.

[mrnorman]

Also include batched reductions to avoid the need for atomics in most of the codes!

[mrnorman]

This Issue is on pause until the need for more functionality arises from MPAS-Ocean folks.

[mrnorman]

Omitted by hierarchical parallelism removal.