fix: refuse dak.Array in hidden-but-public-functions

[jpivarski]

This is solution (3) from #3077 (comment), but I made it a "refuse if dask-awkward," rather than "convert if dask-awkward." They're hidden-but-public (L2) functions, so they can be restrictive. If dask-awkward ever does write overloads for them, those overloads can be a way to avoid this behavior (so we're not making decisions for dask-awkward).

To fix the ATLAS problem, Coffea will need to make an update, to be more careful about dak.Array versus dak.Array._meta.

[jpivarski]

My main question: are there other places in the code like this, which should have a similar dak-check?

I decided to go with type(x).__name__.startswith("dask_awkward."), rather than hasattr(x, "_meta"), since there may be other reasons why a non-dask-awkward array has an attribute named _meta. If class definitions get moved in dask-awkward, then this check should really be

type(x).__name__ == "dask_awkward" or type(x).__name__.startswith("dask_awkward.")

and maybe it should be a utility function.

If there's already a utility function that I just don't know about, please point me to it!