Delta batch-native sink handoff

Applies to: native Delta append and merge-on-read writes

The StreamFusion-owned Delta Sink V2 writer accepts each Arrow output batch as one ownership-carrying record. It walks the batch synchronously with one reusable row cursor for primary-key and RowKind inspection, then stores only row positions for inserts and update-after images that must reach a data file. Update-before and key-only delete records therefore avoid both a StreamRecord and a retained Java object.

Previously the boundary emitted one StreamRecord<RowData> per Arrow row. Delta buffered those wrappers, grouped them by their original batch, and reconstructed columnar selections before the native Parquet write. The current handoff preserves the original Arrow ownership throughout. Dense selections enter the same native encoder as the plain Parquet sink; sparse merge-on-read selections carry row positions into Rust, gather each Arrow column once, and then enter the same standard parquet-rs ArrowWriter. Neither path transposes the data-file payload through Flink rows.

This adapter uses published delta-flink merge and commit APIs. The current released-only 2M-event Nexmark sink diagnostic completed all 23 supported queries at a 1.522× suite geomean over the stock Delta writer; see Benchmarks.