Optimized reduce(vcat, xs) and reduce(hcat, xs) for xs being any iterator

[oxinabox]

We should extend #21672
for all iterators, in so far as is possible.

That way people would not run into surprises
when they do reduce(hcat, filter(x->x[1]>5, xs))
or reduce(vcat, (vec(sum(abs2, xs; dim=2)) for x in xs) etc

Because right now the performance of those kind of things is worse than splatting.
and it is really annoying having to be so careful about your types when writing code.

Likely any PR to solve this would also solve #31169
just because that one makes writing tests for correctness of behavior annoying

Fixing this for generators
would allow solving the mapreduce(vcat/hcat problem in #31137
(and its generalization to nonarrays,)
since
we could have mapreduce(f, vcat, xs) = reduce(vcat, f(x) for x in xs)

[nalimilan]

The difficulty is to avoid ambiguities. I guess you should try and see whether it can work.

[oxinabox]

There are 4 cases:

• vcat of Vectors
• vcat of Matrixs
• hcat of Vectors
• hcat of Matrixs

Potentially, each wanting there own special treatment.
Since for example if you know the length for hcat of Vectors you know the width of the matrix you must allocate.
and any time you work with matrixes reshape tricks need to be played