Add a native Box2D type for bounding boxes

[jiayuasu]

Phase 1 complete. All sub-issues are closed. The Box2D type, its UDT, the SQL surface, cross-language bindings (Python, Flink, Scala DataFrame API), and GeoParquet writer recognition all landed in master.

Summary of what shipped

• Box2D value class + Box2DUDT (struct-backed xmin/ymin/xmax/ymax: double)
• SQL: ST_Box2D, ST_MakeBox2D, ST_Extent (aggregate), ST_GeomFromBox2D, ST_AsText(box2d), accessor overloads ST_XMin/XMax/YMin/YMax(box2d)
• Python bindings (PySpark Box2DType, all wrappers in st_functions / st_constructors / st_aggregates)
• Scala DataFrame API wrappers in st_functions / st_constructors / st_aggregates
• Flink bindings (scalar + aggregate + Box2DTypeSerializer)
• GeoParquet 1.1: Box2D columns are recognized as bbox covering columns in metadata (both explicit geoparquet.covering[.geom]=<col> and auto-detected <geom>_bbox)

Sub-issues

Foundation

• [GH-2877] Add Box2D type and Box2DUDT #2878 — Box2D value class + Box2DUDT (struct-backed) + UDT registration + Python Box2DType

SQL surface

• Implement ST_Box2D(geom) scalar function #2881 — ST_Box2D(geom) scalar
• Overload ST_XMin/XMax/YMin/YMax for Box2D #2882 — ST_XMin / XMax / YMin / YMax(box2d) accessor overloads
• Implement ST_MakeBox2D(p1, p2) scalar constructor #2883 — ST_MakeBox2D(p1, p2) scalar constructor
• Implement ST_Extent aggregate (returns Box2D) #2884 — ST_Extent(geom) aggregate
• Add CAST(box2d AS geometry) and ST_AsText(box2d) #2885 — ST_GeomFromBox2D and ST_AsText(box2d)

Storage

• Emit Box2D columns as GeoParquet 1.1 bbox covering columns #2886 — Box2D columns recognized as GeoParquet 1.1 bbox covering columns in metadata

Cross-language bindings

• Python bindings for Box2D scalar functions and ST_Extent #2887 — Python bindings for the Phase 1 surface
• Flink bindings for Box2D scalar functions and ST_Extent #2888 — Flink bindings for the Phase 1 surface
• Scala/Python DataFrame API wrappers for Box2D Phase 1 functions #2891 — Scala DataFrame API wrappers (folded into [GH-2887] Box2D DataFrame API + Python bindings for Phase 1 #2904 once it surfaced as a hard dependency for the Python bindings)

Deferred follow-ups

These were explicitly scoped out of Phase 1 and will need their own issues if/when prioritized:

• GeoParquet writer: Float32 + conservative outward rounding (Math.nextUp / Math.nextDown). Bit-compatible with apache/sedona-db's next_after. Deferred because writing Float32 today would create a write/read asymmetry — reads come back as struct<float> not Box2D until the reader auto-materialization is in place. Best to land Float32 + reader auto-materialization together.
• GeoParquet reader: auto-materialize covering bbox columns as Box2D when GeoParquet 1.1 metadata points at them. Typed bbox columns from existing files with no migration. The reader path has more edge cases (legacy files, missing metadata, conflicting schemas) — worth its own change.
• ST_Expand(box, dx, dy)
• Box predicates (ST_BoxIntersects, ST_BoxContains)
• Implicit geometry → box2d cast at the Catalyst level
• Box3D, ST_3DExtent, ST_3DMakeBox, ST_ZMin/ZMax — when a concrete user (point clouds, BIM, voxel data) appears
• ST_Box2dFromGeoHash, ST_EstimatedExtent — additional PostGIS box functions
• Geography bboxes — Geography doesn't have a bbox type today; PostGIS doesn't expose one either. Likely path: reuse Box2D with antimeridian-wraparound semantics on the X axis (xmin > xmax), matching apache/sedona-db's WraparoundInterval. This is exactly why we did not burn xmin > xmax on an in-band empty marker.
• R bindings
• Documentation update consolidating the Phase 1 surface in one coherent docs change.

Coordination with sedona-db

sedona-db's GeoParquet writer uses xmin/ymin/xmax/ymax (Float32), but its st_analyze_agg returns minx/miny/maxx/maxy (Float64). Worth aligning on the Parquet-spec naming convention as part of the deferred Float32 writer work above.