feat(rust/sedona-spatial-join): Automatic query-side filter pushdown for KNN joins

[Kontinuation]

Summary

• Adds a KnnQuerySideFilterPushdown optimizer rule that automatically pushes query-side-only filters below the SpatialJoinPlanNode extension node for KNN inner joins
• Only handles INNER JOIN (conservative start); outer join support can be added later
• Updates docs to document the automatic pushdown behavior and clarify when barrier() is still needed

Background

Previously, KNN joins blocked ALL filter pushdown (both query-side and object-side) because the SpatialJoinPlanNode extension node's default prevent_predicate_push_down_columns() returns all columns. Object-side pushdown must remain blocked (it changes KNN candidate sets), but query-side pushdown is safe and should be automatic.

DataFusion's built-in PushDownFilter pushes the same predicate to ALL children of an extension node, so a query-side filter like h.stars >= 4 would fail when applied to the object-side child that doesn't have column h.stars. This requires a custom optimizer rule instead.

Implementation

The KnnQuerySideFilterPushdown rule:

1. Pattern matches Filter(predicate, Extension(SpatialJoinPlanNode)) where the join filter contains ST_KNN
2. Uses find_knn_query_side() to determine which child is the query side (from the first argument of ST_KNN)
3. Splits the filter predicate into conjuncts; pushes query-side-only conjuncts below the extension node; keeps the rest above
4. Runs before DataFusion's PushDownFilter so the pushed-down filters are further optimized into scan nodes in the same pass

Testing

• 12 unit tests for find_st_knn_call and find_knn_query_side
• 3 integration tests verifying correct plan structure (filter pushed into query-side child, object-side filters stay above)

Depends on #635

[paleolimbot]

Thank you!

[paleolimbot]

Optional, but if you have datafusion as a dev dependency already you can also use ctx.parse_sql_expr() to create these (might make it easier to spot errors or create more complex test cases in the future).

[Kontinuation]

It would require registering UDFs with a SessionContext, which adds complexity for what are currently straightforward unit tests. I have replaced these using st_knn_udf.call.

[paleolimbot]

.call() is probably a bit more compact since there are lot of these:

Suggested change

	let knn_expr = Expr::ScalarFunction(ScalarFunction {
	func: st_knn_udf,
	args: vec![col("l.geom"), col("r.geom"), lit(5i32), lit(false)],
	});
	let knn_expr = st_knn_udf.call(vec![col("l.geom"), col("r.geom"), lit(5i32), lit(false)]);

[Kontinuation]

Fixed.

[paleolimbot]

Merging to keep all the PRs in sync since we have a lot open!