Array Constructors - `similar`

I think we are in a good place to start working on array constructors (would also make documenting examples a whole lot easier). This brings up stuff related to similar:

• https://github.com/JuliaLang/julia/issues/25107
• https://github.com/JuliaLang/julia/issues/18161
• https://github.com/JuliaLang/julia/issues/10064

I'm not convinced there's a silver bullet for array constructors, but I think we could at least find a solution to what similar is often trying to do, allow generic method definition without worrying about array specific constructors. I think we actually need to break up what similar does into several pieces though

1. allocate memory - a buffer/Ref that is safer than a pointer
2. do something to the memory - probably fastest on pointers
3. create user facing instance - e.g., Array

I haven't worked out all the details but here's some of the cleaner code I have so far that might support this.

""" allocate_memory(x, args...) """  # step 1 allocates memory with corresponding device to `x` and axes
function allocate_memory(x)  end

""" preserve(f, x, y) """  # step 2 operate on pointers and permit novel garbage collection approaches
preserve(f, x::X) where {X} = preserve(f, x, device(X))
preserve(f, x::X, ::CPUIndex) where {X} = f(x)
function preserve(f, x::X, ::CPUPointer) where {X}
    GC.@preserve x out = f(pointer(x))
    return out
end
preserve(f, x::X, y::Y) where {X,Y} = preserve(p_x -> preserve(Base.Fix1(op, p_x), y), x)


""" as_immutable """
function as_immutable(x) end

""" initialize(original, new_data) """ # step 3 turn processed buffer into user facing array
function initialize(x::X, data) where {X}
    if !ismutable(X)
        return as_immutable(data)
    else
        return data
    end
end

""" instantiate(f, x, allocator, initiializer) """
instantiate(f, x, allocator, initiializer) = initiializer(x, preserve(f, allocator(x),  x))
instantiate(f, x, allocator) = (f, x, allocator, initiialize)
instantiate(f, x) = (f, x, allocate_memory)

The reason I think separating things out like this is helpful is because it turns this function from base like this

function rotr90(A::AbstractMatrix)
    ind1, ind2 = axes(A)
    B = similar(A, (ind2,ind1))
    m = first(ind1)+last(ind1)
    for i=ind1, j=axes(A,2)
        B[j,m-i] = A[i,j]
    end
    return B
end

into this

function rotr90(A)
    ind1, ind2 = axes(A)
    return instantiate(A, x -> allocate_memory(x, (ind2, ind1))) do a, b
        m = first(ind1)+last(ind1)
        for i=ind1, j=axes(A,2)
            b[j,m-i] = a[i,j]
        end
    end
end

This means that new array types typically wouldn't need to change rotr90 but would just change their allocators and initializers. I'm not super familiar with how jagged arrays work but we could have allocate_memory take in device and memory layout info for this.

Start by reading the three linked Julia issues and the proposed allocate_memory, preserve, initialize, and instantiate sketch. The issue names no repository files or tests, and done is not defined beyond reaching agreement on a design for separating array construction responsibilities.