Zero-copy exit transpose

Applies to: the Arrow→RowData exit transpose

What it is

The Arrow→RowData exit transpose used to deep-copy and box every field, because the Arrow batch was closed immediately after conversion. It now emits a reusable lazy ColumnarRowData view over the Arrow vectors instead — the same columnar→row model Spark/Comet use — keeping the batch open through the whole emit loop (5a12f2d).

Why it works

A lazy row view reads fields from the still-open Arrow vectors on demand instead of materializing and boxing every field up front. Combined with Flink object reuse enabled — a standard production setting, applied to both sides of the benchmark — this removes the per-row allocation and copy that used to dominate the exit path.

Measured

Native q0 roughly doubled (713a0a3).