Initialization is broken with limiters on the GPU

[charleskawczynski]

Build here: https://buildkite.com/clima/climaatmos-ci/builds/15369#018c6e50-62d1-46f8-af97-4c987d3b8092/149-222
(this was due to #2270), xref: #2270 (comment). cc @Sbozzolo. Generally we try to avoid UnionAll types like this, that need to be broadcasted, because they're not gpu compatible.

[charleskawczynski]

I'll need to fix this before doing the dt sweep (if we're going to do it on the gpu)

[charleskawczynski]

Maybe the broader issue is that we should add tests for testing that broadcasted structs do not have UnionAll types.

[Sbozzolo]

Can we make the limiter a concrete type instead of UnionAll?

[charleskawczynski]

Can we make the limiter a concrete type instead of UnionAll?

We would need Y to do that, which we don't have until after we've constructed numerics / atmos. For now, I'm proposing, in #2440, that we just use a mirror / singleton type to prescribe if and what limter will be constructed.