Implement reduce_empty(op, Union{})

[tkf]

Interestingly, the invalidation hunt #35733 actually fixed a bug in reduce_empty. As #35733 may not be merged as-is, I extracted the bug fix part from #35733 and added a test to it.

It requires https://github.com/JuliaLang/Statistics.jl/pull/36

[tkf]

@timholy's feedback in JuliaLang/www.julialang.org#794 (comment):

I don't think that's a bug. We have lots of methods that result in an ambiguity error if you use Union{} as a type. Try zero, one, etc.

OK maybe saying it's a "bug" is too much. But I think it is a fault of the caller to invoke one(Union{}). I don't think it's a fault of caller to call sum(Union{}[]) because sum should handle any iterable object and give sensible error when it doesn't work. Union{}[] is a reasonable input for any function on iterables because it is the "zero" of the binary "operator" append!/vcat. So it's the "best" empty vector to use as the initial state for constructing a vector that might contain numbers of unknown type. I don't think revealing the internal gut when a reasonable input is given is ideal.

[tkf]

This also fixes a bug mentioned in #28777 (comment)

On Julia 1.6-DEV:

julia> reduce(*, ())
""

After this PR:

julia> reduce(*, ())
ERROR: ArgumentError: reducing over an empty collection is not allowed

[nalimilan]

Looks good to me. Though note that since (as you showed) reduce(*, ()) == "" currently, making it throw an error may technically break some code.