fix: _getitem_at_placeholder checks

[pfackeldey]

Some implementations of _getitem_at_placeholder have been wrong, e.g. Index classes can't be instances of PlaceholderArrays - only their .data can be.

Also a ListoffsetArray should return true if either its offsets or its contents contain PlaceholderArrays. Now that I think about it, that may be wrong from my side and the check should be only for the offsets (as originally intended). Let me know if this is the case, then I update this PR.

[jpivarski]

I agree 100%—only an Index's data can be a Placeholder, not an Index. Might this have been responsible for some spurious cases of trying to access PlaceholderArrays at runtime? (I haven't followed the logic through, but it seems like it might be related.)

[pfackeldey]

I agree 100%—only an Index's data can be a Placeholder, not an Index. Might this have been responsible for some spurious cases of trying to access PlaceholderArrays at runtime? (I haven't followed the logic through, but it seems like it might be related.)

Unfortunately not likely, unless a user manages to make a repr call be part of the dask graph with dask-awkward - which I doubt. Executing the DAG is non-interactive and there's not really a way to interact with the "materialized" awkward Array during a .compute() phase. _getitem_at_placeholder() is only used for constructing the repr of an awkward Array that may contain PlaceholderArrays as far as I understand.

Without this PR the repr of awkward Arrays that contain ListoffsetArrays (and some others that have Index instances on them) would lead to an error; I stumbled over this while playing around with dask-contrib/dask-awkward#565.

[ianna]

@pfackeldey - Great, thanks! Please, merge it if you finished with it. Thanks!