throw if no fields

[aplavin]

Before, StructArray([nothing]) or StructArray([1, 2, 3]) created empty (zero-length) arrays. That was silent and pretty confusing.

[aplavin]

bump

[piever]

Should this be considered breaking? In that case, may it would be best to coordinate merging this one with other breaking changes.

[aplavin]

As I understand, the StructArray construction for types without fields just silently returned a wrong and totally unexpected result. This PR makes it an error.
Do you think the previous behavior was somehow useful in certain cases?

The ideal would be to actually support such types, but I don't see a simple place to plug that support in.

[piever]

Do you think the previous behavior was somehow useful in certain cases?

I think IndexedTables used the "StructArray with no fields" in some cases (cc: @shashi), which is why the constructor is there in the first place, but I confess I don't remember the exact usecase.

[aplavin]

Took all comments into account in the recent commit.

Also see #242, alternative to this PR. Here I make it throw an error instead of a silently wrong behavior, and #242 makes no-field eltypes actually work.

[piever]

Thanks for finalizing this. I think I'd prefer this over #242. It might make sense to combine this and #243 (which also seems very slightly breaking) into the next breaking release.

[aplavin]

Great, looking forward for this to be released!
Is this PR really breaking? It looks strictly as a bugfix to me: unsupported arguments throw an error instead of returning an obviously wrong result.
As for #243, technically it could break someones code with very strict component type checks, but these things are regularly considered non-breaking even by base julia.

There were multiple changes recently that are arguably more breaking, but didn't get a breaking release.
Take #239 for example: if someone uses StaticArrays <= 1.5.3, they'll have StructArrays not working with FieldArrays properly.
Or #227 breaking some fancy array types, as discussed in #228.
Almost any change is technically breaking for some.
Maybe, a good indicator for usefully breaking/non-breaking release is whether any existing tests were modified?

Anyway, a breaking release would also be fine for me.