Add check for 0-dimensional array arguments in hvncat and produce an error

[BioTurboNick]

Resolves #41047 by checking for zero-length inputs, accounting for the case where the shape form of hvncat has higher-dimensional arguments than the highest dimension of concatenation, and appropriate tests for all.

[mbauman]

This is outside the commentable diff, but:

julia/base/abstractarray.jl

Line 2295 in de1444c

nd = max(N, cat_ndims(as[1]))

Is it possible for as to be empty?

[DilumAluthge]

@mbauman CI is green now. Is this good to merge?

[JeffBezanson]

I don't see why this is a special case, i.e. why isn't it covered by the other check that the number of elements matches?

[JeffBezanson]

For example [zeros(1,2,0);;;7 8] doesn't need to be an error.

[BioTurboNick]

The former check is the length of as; this is the length of each element. Though I should probably be using cat_length here in case any of them are strings.

But you're right. This is a crude check. The dimension calculating algorithm right now just can't handle arrays with a zero dimension. e.g. [zeros(1, 0);;;7 8] creates an underfilled array. Seems better to make them error for now until I can address that.

[BioTurboNick]

I have a better solution, which will end up in #41143, that allows correct zero-length arrays to work the way you would expect. So this PR can be closed.