Allow type-unstable functions in mapslices

[ararslan]

Here's my attempt at fixing #26232, which doubles as an attempt to convince someone else to fix it better.

[nalimilan]

I'm afraid the real fix is going to be more complex: we should use the same approach as map, i.e. reallocate the array if we encounter a slice with a new type.

[martinholters]

Yep, we shouldn't use return_types to determine the element type of non-empty containers.

[mbauman]

Oh I hate mapslices with a passion. Let's kill it.

In your use-case, is the inner function returning a scalar or an array? If it's a scalar, then use map(f, eachslice(A, dims=d)) — that'll "just work." If you need support for multiple dims, let's focus our efforts on adding that capability to eachslice.

If you're returning an array, though, map eachslice will return an array of arrays. You can then concatenate them together yourself, but there's definitely a performance advantage to fusing the two together. What if we were to internally use a datastructure like (view(R, :, i) for i in I) that would expose columns of a matrix as though they were single elements? And then used map!? That's not quite right, but I'd like to pursue something along these lines to pave the way towards a more consistent, reimagined version of this nonscalar application of mapslices.

[ararslan]

See, my plan of getting someone else to fix it better is already working 😄

[ararslan]

I will not be the one to fix it better. Hopefully someone else will. :)