Do we need separate MultiCPU and MultiGPU architectures?

[glwagner]

The two structs are identical:

Oceananigans.jl/src/Distributed/multi_architectures.jl

Lines 6 to 22 in a3faff7

	struct MultiCPU{G, R, I, ρ, C, γ} <: AbstractCPUArchitecture
	distributed_grid :: G
	local_rank :: R
	local_index :: I
	ranks :: ρ
	connectivity :: C
	communicator :: γ
	end
	struct MultiGPU{G, R, I, ρ, C, γ} <: AbstractGPUArchitecture
	distributed_grid :: G
	local_rank :: R
	local_index :: I
	ranks :: ρ
	connectivity :: C
	communicator :: γ
	end

An alternate design would use a single DistributedArchitecture (or something), with a property child_architecture that could be CPU or GPU.

Doing this would require us to get rid of AbstractCPUArchitecture and AbstractGPUArchitecture (or turn them into type unions). I don't think we use these abstract types, do we?

@ali-ramadhan @kpamnany

[glwagner]

This are a bit different now