column on grids are inconsistent and do not actually make views

[charleskawczynski]

#1487 fixed some issues, but column does not actually make a view, so the underlying DataLayout for a VIJFH-backed ExtrudedFiniteDifferenceSpace is still a VIJFH DataLayout.

The issue here is that column on a broadcasted object now contains a mixture of VF and VIJFH datalayouts. Memory access patterns for our metric terms are now "stuck" to those governed by VIJFH. i.e., we cannot make MArray-backed datalayouts.

This is a hard block, stopping #1763 because we cannot make an MArray-backed VIJFH (as that would be huge), and we cannot reconstruct ExtrudedFiniteDifferenceGrid on the GPU, since it is mutable.

What we have now seems like a bad design in that I would expect column to recurse all the way down a broadcasted object, and it doesn't. It's also unfortunate that this is blocking #1763.

The same is probably true for slab.

[charleskawczynski]

Here's a reproducer that can be used to help see what's going on:

import ClimaCore, ClimaComms, CUDA
@isdefined(TU) || include(
    joinpath(pkgdir(ClimaCore), "test", "TestUtilities", "TestUtilities.jl"),
);
import .TestUtilities as TU;
using Test, StaticArrays, IntervalSets, LinearAlgebra, BenchmarkTools
import ClimaCore: Domains, Spaces, Fields, Operators
import ClimaCore.Domains: Geometry
import LazyBroadcast as LB

const CT3 = Geometry.Contravariant3Vector
const C12 = ClimaCore.Geometry.Covariant12Vector
const ᶠwinterp = Operators.WeightedInterpolateC2F(
    bottom = Operators.Extrapolate(),
    top = Operators.Extrapolate(),
);
FT = Float64;
cspace = TU.CenterExtrudedFiniteDifferenceSpace(FT; zelem = 25, helem = 10);
fspace = Spaces.FaceExtrudedFiniteDifferenceSpace(cspace);
ᶜx = fill((; uₕ = zero(C12{FT}), ρ = FT(0)), cspace);
ᶠx = fill((; ᶠuₕ³ = zero(CT3{FT})), fspace);
ᶜJ = Fields.local_geometry_field(ᶜx).J;
bc = LB.@lazy_broadcasted @. ᶠx.ᶠuₕ³ = ᶠwinterp(ᶜx.ρ * ᶜJ, CT3(ᶜx.uₕ));

bc_col = Fields.column(bc, 1,1,1)