nested broadcast should evaluate non-broadcasted part once

[tpapp]

MWE following discussion:

struct Bar
    data
    Bar(data) = (println("expensive"); new(data))
end
(B::Bar)(x) = B.data * x
Bar(2).([1,2,3])                  # Bar(2) will be only evaluated once
broadcast(sqrt, Bar(2).([1,2,3])) # Bar(2) will also be evaluated once
sqrt.(Bar(2).([1,2,3]))           # Bar(2) will be evaluated three times

[nalimilan]

Making for triage as changing this to evaluate Bar(2) only once is technically breaking, though probably few packages rely on it (so it could be fixed after the alpha).

[StefanKarpinski]

This would be good, but it seems like some effort. I would propose documenting that this may change and then changing it in 1.x at some point. As already noted, it's unlikely that people are relying on this.

[vtjnash]

Should be fixed by #23692, although Jeff doesn't consider this a bug, so it can get changed anytime

[JeffBezanson]

In these kinds of functional programming contexts it's not a good idea to depend on side effects happening in a particular way. For example, does map on a sparse matrix call the function again for every non-stored zero?

Yes, I consider this mostly undefined behavior, so it's fine to change it as well.