Notable changes:

• We now use fine for encoding/decoding throughout. In particular for structs such as %Column{}, %Field{}, %BufferData{}.
• We use exceptions to propagate errors while building/reading arrow data, instead of status returns.
• Using nanoarrow::UniqueSchema, nanoarrow::UniqueArray for RAII memory management, which makes it easier to avoid memory leaks.
• I removed materialization, so now we build %Adbc.Column{} with both :field and :data in a single pass. Recently we switched materialization, so that it doesn't build individual terms in the NIF, but instead we return simple structs like %BufferData{} with resource binary inside. This means that the more expensive part is not materialization, but decoding later as part of to_list.