Fix Adapt calls, and remove device-side structs

[charleskawczynski]

In efforts to support conversion between cpu and gpu, we need the device-side objects to retain as much as possible so that we can convert back and forth without losing information.

This PR does a few things:

• Abstracts away getproperty calls to mpi context for mpicomm
• Fixes some adapt calls-- we only want adapt_structure to return device structs when called with a CUDA.KernelAdaptor.
• Remove the device side structs, and instead use the ClimaComms objects.

This should hopefully work for all cases except for CUDA-MPI jobs, since adapting from CPU to GPU with mpicomm (a mutable object) cannot reside on the device. This limitation should hopefully be fixed by CliMA/ClimaComms.jl#102.

Also, I noticed that the main branch has some typing issues. In particular:

## aliases
const RectilinearSpectralElementGrid2D =
    SpectralElementGrid2D{<:Topologies.RectilinearTopology2D}

does not apply to device-side objects, since those would need to be

const RectilinearSpectralElementGrid2D =
    DeviceSpectralElementGrid2D{<:Topologies.RectilinearTopology2D}

(assuming that their type parameters are synchronized)

[sriharshakandala]

Would we be able to

1. Move all Adapt code out of src/ to ext/?
2. Avoid adding Adapt explicitly to the source project and extensions as a dependency, instead access it through CUDA.Adapt or Metal.Adapt inside the extensions?
   For example, something similar to https://github.com/CliMA/ClimaInterpolations.jl/pull/3/files#diff-3702a35dcf2bad10ba7fc72fc0c50366627cd0679c4e3afda9d24c602ad82a8eR117 &

[charleskawczynski]

Would we be able to

1. Move all Adapt code out of src/ to ext/?
2. Avoid adding Adapt explicitly to the source project and extensions as a dependency, instead access it through CUDA.Adapt or Metal.Adapt inside the extensions?
   For example, something similar to https://github.com/CliMA/ClimaInterpolations.jl/pull/3/files#diff-3702a35dcf2bad10ba7fc72fc0c50366627cd0679c4e3afda9d24c602ad82a8eR117 &

I think we should be able to move adapt code into ext/, if not all of them, but there may be some cases where we need cpu adapt code.

In particular, there may be a difference between cuda adapting for the purposes of prepping data for kernels, and us adapting data between the cpu and gpu.

[charleskawczynski]

Actually, we can overload getproperty on the ClimaComms context, and make it backwards compatible. So we don't even need to change the .mpicomm calls here. Going to revert that.

[charleskawczynski]

Closing in favor of #2120 and #2114.