Make CUDA an extension

[charleskawczynski]

This PR makes CUDA an extension.

This should allow us to more easily test on newer versions, if cuda is a bit late to catching up. Also, it should help with code loading times when we're only running on the cpu.

Closes: #1667

[charleskawczynski]

A couple quarks I applied are, so that we can overload them with cuda:

• cuda_synchronize(device::ClimaComms.AbstractDevice; kwargs...) = nothing
• allow_scalar_func(::ClimaComms.AbstractDevice, f, args) = f(args...) & allow_scalar_func(::ClimaComms.CUDADevice, f, args) = CUDA.@allowscalar f(args...)
• is_CuArray_type(::Type{T}) where {T <: CUDA.CuArray} = true

I had to move a few remapping interpolate field functions, so that we can dispatch between device types.