Make PyArrow Dependency Optional

[WillAyd]

Needs some investigation, but I think we have a feasible path to replacing pyarrow with arro3 internally. The only thing we use pyarrow for is to create a recordbatchreader and convert that into the appropriate end dataframe libraries.

If we can replace that with arro3, it should save a good deal of installation size

[WillAyd]

Actually we don't need arro3 or pyarrow for cases where users opt for the capsule return type implemented in #378

If we dropped pyarrow as a dependency, we might just have to add a check in the reader like:

if return_type != "stream":
    import pyarrow as pa
    ... # handle error if not installed