Minor optimization: use convert rather than constructor to convert array type before output

[glwagner]

Currently, we use the constructor output_writer.array_type to convert array data prior to outputting:

Oceananigans.jl/src/OutputWriters/fetch_output.jl

Line 17 in 03a6f85

convert_output(output::AbstractArray, writer) = CUDA.@allowscalar writer.array_type(output)

But we should use convert instead, because this avoids allocating memory when no type conversion is needed:

julia> a = rand(1, 1, 1)
1×1×1 Array{Float64,3}:
[:, :, 1] =
 0.7727202498256802

julia> b = convert(Array{Float64}, a)
1×1×1 Array{Float64,3}:
[:, :, 1] =
 0.7727202498256802

julia> b === a
true

so b is just a reference to a, but

julia> c = Array{Float64}(a)
1×1×1 Array{Float64,3}:
[:, :, 1] =
 0.7727202498256802

julia> c === a
false

c is not a reference to a.

This matters very little since we basically always need to allocate memory to convert from Float64 to Float32.

A related minor optimization would be to avoid converting views to the same type as the parent array, since we could just output those directly. I think we have to use dispatch on one of the type parameters of SubArray for that. But maybe simpler code is worth not implementing a minor optimization for an edge case.

[glwagner]

I'm closing this issue because I'm judging that it's not of current, timely relevance to Oceananigans development. If you would like to make it a higher priority or if you think the issue was closed in error please feel free to re-open.