Projection and filter operators (Volcano model)

Add two physical operators on top of your scan, in the classic Volcano / iterator style: each operator pulls RecordBatches from its child and pushes transformed batches to its parent.

Task

Make execute() handle these plan nodes:

• Project(child, columns) — select/reorder columns.
• Filter(child, predicate) — keep rows where predicate is true.

Predicates are expression trees over the stage-1 API:

Filter(scan, Compare(Col("amt"), ">", Lit(25.0)))
Filter(scan, And(Compare(Col("amt"), ">=", Lit(10.0)),
                 Or(Compare(Col("cat"), "=", Lit("a")),
                    Compare(Col("cat"), "=", Lit("b")))))

Requirements

• Evaluate predicates with pyarrow.compute: map each Compare.op to the right kernel (equal, less, greater, …) and combine And/Or with and_kleene / or_kleene so NULLs follow SQL three-valued logic.
• Project and Filter must compose with each other and with Scan.
• Results match DuckDB for the same WHERE / column list (G6). A Hypothesis property checks filter soundness: every surviving row satisfies the predicate and the row count matches the oracle.

Constraints

• Stay in the iterator model — operators pull batches, they don't re-read files.
• Earlier stage (scan) must keep passing: de test runs stages 1–2 together.