These changes get kokkos+rrtmgp performance on frontier to match yakl+rrtmgp

[jgfouca]

Performance was measured with this case:
SMS_Ln300.ne30pg2_ne30pg2.F2010-SCREAMv1.frontier-scream-gpu_crayclang-scream.scream-perf_test--scream-output-preset-1

With these changes, we appear to be spending less time in kokkos kernels than yakl kernels:

Total time spent in YAKL kernels:   16.262377
Total time spent in Kokkos kernels: 14.596052

Change list:

• Replaces all multi-dimensional kernel launches with macro FLATTEN_MD_KERNEL$N. This removes most uses of Kokkos' MDRange policy, which does not seem to perform as well in most cases.
• YAKL SimpleBounds always has the right index being the fast one; it was inverting the dimensions to get layout left, ex:

parallel_for( YAKL_AUTO_LABEL() , SimpleBounds<3>(a,b,c) {
   md_array(c, b, a)
}

^ Notice the index order flip. FLATTEN_MD_KERNEL works for both left and right layouts and will do the right thing in either case, so there's no need to do this inversion..

• Get rid of all uses of alloc_raw to allocate multiple temporary views at the same time. This was leading to views that were not cache aligned which hurt performance.
• Adds non-allocating versions of some routines. YAKL didn't have to worry about this since it was doing pool allocations for all arrays automatically. The non-allocating routines take a view that has already been created (presumably via the pool allocator).
• Add fences to timing macros to ensure accurate times
• Tweaks to the pool allocator to improve performance. Round up all allocations to cache line size.
• Prefer type_check_v<type> over type_check<type>::value

[jgfouca]

Performance now seems to match YAKL on machines we care about. This is ready for final review.

[ambrad]

You should wait for Ben's approval, but this looks good to me. One thing that would be useful: Put a comment in somewhere explaining the naming scheme for "init_no_alloc" and "alloc_no_alloc".

[jgfouca]

Some of the interesting changes are here. These are the functions that unflatten an MD idx and the macros that makes this stuff usable.

[jgfouca]

Interesting code here. How temp views are allocated definitely impacts performance.

[jgfouca]

@ambrad , you beat me to the punch! I left a few annotations for the parts that I think are interesting.

[jgfouca]

Oh, I also added more to the PR description.