Fix dispatching for some cuda kernels

[charleskawczynski]

This should close #1786, but I think we need more careful testing around this.

This PR:

• revamps and adds tests for fill!
• Makes size return the correct number of fields for all datalayouts, instead of returning 1 (this may need to be reverted if we rely on this).
• Fixes a bug in our setindex! for IFH datalayouts.

cc @trontrytel

[charleskawczynski]

So, it turns out that I discovered a rather insidious bug in our DataLayouts. Both of our setindex! methods in our IFH DataLayouts use Val(3) (wrong) instead of Val(2) (correct) for the field index:

@inline function Base.setindex!(data::IFH{S}, val, i, _, _, _, h) where {S}
    @inbounds set_struct!(
        parent(data),
        convert(S, val),
        Val(3),
        CartesianIndex(i, 1, h),
    )
end
@inline function Base.setindex!(
    data::IFH{S},
    val,
    I::CartesianIndex{5},
) where {S}
    i, _, _, _, h = I.I
    @inbounds set_struct!(
        parent(data),
        convert(S, val),
        Val(3),
        CartesianIndex(i, 1, h),
    )
end

The (very simple) test that I added for fill is what caught this. I looked at all of the other datalayouts, and it seems like this was the only place where we made this mistake.

This impacts both the CPU and gpu. Looking around at ClimaCore internals, this may impact:

• when calling level on VIFH datalayouts, and following logic
• limiters (1D/2D cases)
• SpectralElementGrids with GLL quadratures

[charleskawczynski]

Ok, CI is currently broken:

• The latest SciMLBase is now calling ndims(::Type{<:Field})/ndims(::Type{<:AbstractData}), so we need to define it.
• Also, the docs are not running with the latest version of ClimaCore

I've also joined a few docs from DataLayouts, because we cannot document a struct and outer constructors separately (at least Documenter doesn't seem to be picking it up).

There's a lot more that we can do, but I'd like to split off additional work so that this PR doesn't spiral out of control.

[charleskawczynski]

Alright, I think that this should (now) work..

[charleskawczynski]

Alright, the only failure is due to a fix needed in multibroadcast field copyto!s, I just fixed and checked it locally, I'm going to manually merge instead of re-running CI.

[Sbozzolo]

Alright, the only failure is due to a fix needed in multibroadcast field copyto!s, I just fixed and checked it locally, I'm going to manually merge instead of re-running CI.

Okay, I'd like to merge #1788 next (once we find a better name), so maybe you can add a commit to that PR that bumps the minimum version of MBF with the fix, once you release it.

[charleskawczynski]

Okay, I'd like to merge #1788 next (once we find a better name), so maybe you can add a commit to that PR that bumps the minimum version of MBF with the fix, once you release it.

It's already done, the fix was already incorporated in this commit

[Sbozzolo]

Perfect, thank you!