Bounded Arrow batch sizing

Bounded pipelines prefer 8,192 rows per Arrow batch, matching Comet's default. The entry transpose and bounded DataFusion operators capture that preference at planning time. Explicit streamfusion.transpose.batchRows table/JVM settings take precedence; streaming defaults are unchanged. This amortizes JNI, input copying, Arrow IPC framing and batch-level state updates without removing conversions or adding memory authority. Working state remains within the existing shared cap.

TPC-H Q3 over identical stock Parquet files initially regressed at 1,024 rows (1,152.00 ms stock, 1,351.58 ms native). After profiling, 8,192-row batches gave 1,254.80 ms stock versus 1,029.47 ms native, a 1.219× complete-job gain. Both cohorts use scale factor one, parallelism two, the same memory limits, three warmups and seven alternating trials. Both perimeter transposes, shuffles, scans, JNI and completed-result consumption remain timed. These shared-host cohorts have different stock medians, so retain the raw comparisons and assess each native result against its matching stock trials. This is a per-query improvement, not a full TPC-H acceleration claim.