feat(python/sedonadb-geopandas): add sjoin, dissolve, column assignment, and arithmetic

[jiayuasu]

Extends the experimental sedonadb-geopandas package (#1052) with the operations a real pipeline needs: a spatial join, a dissolve, column assignment, and arithmetic on columns.

gdf = sgpd.from_geopandas(points)
gdf["density"] = gdf["pop"] / gdf["area"]          # assign a computed column
joined = gdf.sjoin(regions, predicate="within")    # spatial join
zones = joined.dissolve(by="region")               # group and union geometry

What is added

• sjoin(other, how, predicate, lsuffix, rsuffix, distance) — how of inner, left, or right, over the intersects, within, contains, touches, crosses, overlaps, covers, covered_by, and dwithin predicates. The predicate reads left-relative-to-right as in GeoPandas. The result carries exactly one geometry column — the left frame's, or the right frame's for how="right", matching which side GeoPandas keeps — and column names occurring on both sides get the lsuffix/rsuffix treatment.
• dissolve(by, aggfunc="first") — unions each group's geometry via ST_Union_Agg and carries the remaining columns. by=None dissolves everything into one row.
• __setitem__ — gdf["x"] = ... from a Series of the same frame, a SedonaDB expression, or a scalar to broadcast.
• Arithmetic on Series — +, -, *, /, unary -, and the reflected forms.
• Series.expr — the underlying expression, as an escape hatch for anything the wrapper does not cover (series.expr.funcs.st_point(other.expr)), assignable back onto a frame.

Two semantics worth review attention

/ follows pandas, not SQL. The engine truncates when dividing integers, so v / 2 over an integer column would return floored values where pandas returns 0.5, 1.5, and so on. Since the point of this package is that GeoPandas code keeps working, the numerator is cast to double first. // is deliberately not implemented rather than mapped onto SQL division, which truncates toward zero where Python floors — a subtly wrong operator seemed worse than a missing one.

Assignment rebinds the frame. The underlying frame is immutable, so gdf["x"] = ... replaces the frame this object points at. A Series captured before that assignment therefore belongs to the previous frame, and combining it with a later read raises rather than silently mixing two frames. The error message says so explicitly.

Deviations from GeoPandas, all documented in the README

• sjoin produces no index_left/index_right column, and dissolve leaves the group keys as ordinary columns rather than moving them into the index. Both follow from there being no row index.
• Columns cannot be combined across two frames; without row alignment that would be a silent wrong answer, so it raises and suggests joining first.

Verification

Checked against GeoPandas rather than only for self-consistency:

• sjoin column shapes match exactly for all three how values (after dropping the index_* columns), as do the active geometry column and the row count.
• sjoin row pairings match.
• dissolve areas and first-aggregated column values match ([7.0, 4.0] and [1, 3]).
• Arithmetic matches, including the reflected forms.

46 tests pass locally and in a clean virtual environment installing sedonadb from the nightly index — the environment shape that caught the missing dependency in #1134. ruff format and ruff check are clean.

Deferred

The GeoSeries accessor batch (geom_type, x/y, bounds, envelope, boundary, to_wkt, and friends) is left for a follow-up. One compatibility wrinkle to settle there: ST_GeometryType returns ST_Polygon where GeoPandas' geom_type returns Polygon, so that mapping needs a decision rather than a quick pass-through.