[BUG] make_empty_column produces a dictionary column in an invalid state

[vyasr]

Describe the bug
Currently running the following two lines will produce a segmentation fault:

  auto col = cudf::make_empty_column(cudf::data_type{cudf::type_id::DICTIONARY32});
  auto x = cudf::dictionary_column_view(col->view()).indices();

The reason for this is that when producing an empty column the make_empty_column function does not produce any children for the column, resulting in a null pointer dereference on the indices() call.

In principle, the function protects against this class of error by explicitly opting out of supporting nested column types. However, dictionary types are not considered nested by the is_nested trait.

Expected behavior
This code should either succeed or fail gracefully, not seg fault.

Additional context
The two obvious options for solutions would be to add dictionary types to is_nested or to manually disable them in the factory function. Naively I would opt for the former, but I suspect that would violate expectations in other parts of the code. I unfortunately don't know the dictionary code very well.

[vyasr]

@davidwendt I'm happy to open a PR with a fix if you can advise on what makes the most sense here.

[davidwendt]

Should I be concerned that we are suddenly trying to use dictionary types?

Accessing children of an empty compound column (one with potential children) is considered UB.
https://github.com/rapidsai/cudf/blob/branch-25.04/cpp/doxygen/developer_guide/DEVELOPER_GUIDE.md#empty-columns
I can explicitly add dictionary types this section of the developer's guide if necessary.

The following will also segfault:

  auto col = cudf::make_empty_column(cudf::data_type{cudf::type_id::STRING});
  auto x = cudf::strings_column_view(col->view()).offsets();

[vyasr]

No, nothing to be worried about here. This came up in #18084 because we already support dictionary type in interop. This discussion implies that we have UB here. I can fix that in a follow-up PR.

[davidwendt]

Attempting to fix this here #18121

[vyasr]

Thank you!