Update to PyArrow v21 and native GeoParquet outputs

[zaneselvans]

Overview

• As of v21.0.0 Arrow (and PyArrow) support the GEOMETRY and GEOGRAPHY spatial types natively via Arrow extension types for geoarrow.wkb (issue: [Parquet][C++] Implement Geography and Geometry types in the C++ Parquet implementation apache/arrow#45522 & PR: GH-45522: [Parquet][C++] Parquet GEOMETRY and GEOGRAPHY logical type implementations apache/arrow#45459
• Hopefully this means that we no longer need to do the special munging of geospatial + categorial types to construct the schema for our Parquet/GeoParquet IO Manager -- it should all work within generic PyArrow/Parquet once we've updated to v21.
• However, these outputs may or may not be compatible with older tooling. For example, Kaggle's python environment is still pinned to the now quite stale GeoPandas 0.14 (from April, 2024).

Success Criteria

• We have consolidated our Parquet & GeoParquet IO Managers into a single generic Parquet IO Manager that can handle both spatial and non-spatial outputs.
• Our spatial and non-spatial Parquet outputs are easy to read in with Pandas/GeoPandas, DuckDB (with Spatial extensions), R, and Polars.
• Our spatial and non-spatial Parquet outputs still work with our Kaggle notebooks and their python environment... or we have moved our notebooks to another setup where we can specify a more up to date python environment.

[jdangerx]

@zaneselvans this is blocked until Kaggle's environment is updated, right?

[zaneselvans]

@jdangerx I'm not sure. I'm suspicious that there's something about my attempt to add the geospatial metadata explicitly, and the now official native support being in conflict with each other that is what's causing the older geopandas version on Kaggle to fail, and I am wondering if we just switched over to using the real native support if it would all start working smoothly.