Find a better solution for determining element type for the return array in e.g. normalize

[BioTurboNick]

Coming from: JuliaLang/julia#55830

Making this an issue rather than PR because there seems to be some question about what the solution should actually be, and it might be broader than this one method.

In the normalize stack, the existing code computes the result on one element, finds the type, and then that becomes the element type for the output array. I believe this is done to be generic for operations on arrays of arrays. However, relying on the first element means the type of the result has a hidden dependency on element order, and the scalar lookup means the code can't run on a GPU.