RFC: Add similar for PermutedDimsArray

[Keno]

This allows you to do things like:

julia> A = [1 2 3;4 5 6]
2×3 Array{Int64,2}:
 1  2  3
 4  5  6

julia> X = PermutedDimsArray(A, [2,1])
3×2 PermutedDimsArray{Int64,2,(2, 1),(2, 1),Array{Int64,2}}:
 1  4
 2  5
 3  6

julia> Y = Base.collect_similar(X, (x for x in Any[1 2.0; 3 4; 4 6]))
3×2 PermutedDimsArray{Real,2,(2, 1),(2, 1),Array{Real,2}}:
 1  2.0
 3  4
 4  6

julia> Y.parent
2×3 Array{Real,2}:
 1    3  4
 2.0  4  6

In particular I wanted for collecting into an array from an iterator that yielded its elements in transposed order, but it seems like a sensible definition regardless.

[timholy]

Good idea

[timholy]

Do you need to define any more than this one definition? (Do fallbacks not handle the rest?)

[timholy]

Is this inferrable? (Do we need ntuple(f, Val{N})?)

[andyferris]

Hmm... here was about to use PermutedDimsArray to do lazy transpose matrices and not have this behavior... :)

Is there a convenient way to get a "similar" array that strips the PermutedDims wrapper?

[timholy]

Is there a convenient way to get a "similar" array that strips the PermutedDims wrapper?

Does similar(Array{eltype(A)}, indices(A)) do what you want? (You can specialize this syntax, of course.)

[andyferris]

Well, it would work when you want Array... I'm looking for similar of the parent.

What I really want (and will implement separately to PermutedDimsArray if required) is similar(m::TransposedMatrix, T::Type, d::Dims) = similar(parent(m), T, d). I was just about to hook into const TransposedMatrix = PermutedDimsArray{(2,1)} (shortened form).

None of this means this PR is bad. But when I implement similar on my own generic wrapper types I tend to ditch the wrapper (since they tend to add a layer of inefficiency) and use the parent to control similar. The minimal PR for that here is:

function Base.similar(a::PermutedDimsArray, ::Type{T}, dims::Dims) where {T}
    similar(a.parent, T, dims)
end

[andreasnoack]

Maybe time to start implementing the storagetype function that we have been discussing the PRs to remove full?