fix(duckdb): support materializing enum types to pyarrow

[NickCrews]

See the added test, and the comment.

I'm not sure if this approach, of pa_table.cast(ibis_table.schema().to_pyarrow()), is too broad.
Instead of casting the entire pyarrow table to the expected schema, we could look for the specific case of
when the ibis type is string and the pyarrow type is dictionary<values=string, indices=[intlike]>), and just convert those columns. For instance, this approach could be masking some other semantic mistake that we are making. I assume that pyarrow is smart enough to not actually perform computation unless the cast is really required, so I don't think this should be computationally wasteful.

This also makes it much more explicit that to get to pandas, we first convert to pyarrow, then to pandas, further cementing pyarrow as our first-class citizen, before pandas. I think I remember seeing this as a desire in a different issue/comment. It certainly makes it easier to reason about consistencies between to_pyarrow() and to_pandas()