Unflatten - Implementing views instead of new vectors

[paschermayr]

From: #39

This only relates to array valued parameter. At the moment, when a Vector is unflattened, a new Vector is created for each argument in the tuple:

function flatten(::Type{T}, x::Tuple) where {T<:Real}
    x_vecs_and_backs = map(val -> flatten(T, val), x)
    x_vecs, x_backs = first.(x_vecs_and_backs), last.(x_vecs_and_backs)
    lengths = map(length, x_vecs)
    sz = _cumsum(lengths)
    function unflatten_to_Tuple(v::AbstractVector{<:Real})
        map(x_backs, lengths, sz) do x_back, l, s
            return x_back(v[(s - l + 1):s]) #HERE
        end
    end
    return reduce(vcat, x_vecs), unflatten_to_Tuple
end

This is necessary, as otherwise the unflattened NamedTuple would contain a bunch of Subarrays, as we cannot deduce the original array from

function flatten(::Type{T}, x::NamedTuple{names}) where {T<:Real,names}
    x_vec, unflatten = flatten(T, values(x))
    function unflatten_to_NamedTuple(v::AbstractVector{<:Real})
        v_vec_vec = unflatten(v)
        return NamedTuple{names}(v_vec_vec) #HERE
    end
    return x_vec, unflatten_to_NamedTuple
end

Changing return NamedTuple{names}(v_vec_vec) to typeof(x)(v_vec_vec) would allow us to use views in the first code block, but this change would make many AD backends unusable.