Zero-copy batch handover across JNI

Applies to: every batch crossing the Java↔Rust boundary

Batches cross the Java↔Rust boundary via the Arrow C Data Interface — the consumer reads the producer's memory in place, no serialization, with ownership transferred so buffers are released exactly once.

Batching amortizes the boundary

Rows are buffered into batches before crossing into native code, so the JNI cost is paid once per thousands of records instead of per record. This is what makes the crossing pay off at all versus per-row execution.