Clarify inlining of unrolled_reduce for non-orographic gravity wave

[dennisYatunin]

This PR replaces the call to unrolled_reduce(op, Val(nc), init) in the non-orographic gravity wave parametrization with a forcibly non-inlined call to unrolled_reduce(op, StaticOneTo(nc), init).

The original solution was introduced in CliMA/UnrolledUtilities.jl#10 and CliMA/UnrolledUtilities.jl#12 in order to prevent compilation of 3 unit tests in CI from hanging (nogw_test_3d.jl, nogw_test_mima.jl, and nogw_test_single_column.jl), but it does not have a direct equivalent in the standard library. The new version of this unrolled function call, which achieves the same goal of preventing compilation from hanging, is exactly equivalent to reduce(op, Base.OneTo(nc); init) (and also to reduce(op, StaticOneTo(nc); init)). The new version also makes it clear that inlining of the unrolled function call is what causes compilation to hang in this case.

Note 1: This does not constitute a general rule of thumb, as it is possible to construct similar examples where not inlining is what causes compilation to hang. For example, if itr is a StaticBitVector that contains 256 Bools, then the following code hangs without inlining of unrolled_reduce:

unrolled_reduce(((itr′, i)->unrolled_reduce(((itr′′, j)->unrolled_reduce(((itr′′′, k)->Base.setindex(itr′′′, !(itr′′′[min(i, j, k)]), k)), StaticOneTo(length(itr′′)); init = itr′′)), StaticOneTo(length(itr′)); init = itr′)), StaticOneTo(length(itr)); init = itr)

Inlining has been observed to improve compilation of several unit tests in UnrolledUtilities, whereas not inlining has only been observed to improve compilation of this one particular example in ClimaAtmos. So, UnrolledUtilities will continue to inline everything by default, and @noinline can be used on a case-by-case basis when compilation is too slow.

Note 2: Although the equivalent non-unrolled function call uses init as a keyword argument, and although unrolled_reduce supports using init as a keyword argument, this is not done here in order to ensure GPU compatibility. If the non-orographic gravity wave parameterization calls unrolled_reduce(op, StaticOneTo(nc); init), the GPU compiler claims that it cannot compile the resulting kernel because it contains a dynamic function call.

@Xiaoyang-Xie, the compilation mystery is finally resolved!

[dennisYatunin]

Looks like CliMA/UnrolledUtilities.jl#13 didn't help with the compilation hanging issue. I will now try out CliMA/UnrolledUtilities.jl#15 instead.

[dennisYatunin]

CliMA/UnrolledUtilities.jl#15 also didn't fix the issue. Now trying out CliMA/UnrolledUtilities.jl#16. We might need to combine these two PRs before we have a solution, but maybe this will work on its own.