[VL] Shuffle write size larger than vanilla spark

### Description

Conf:
spark.gluten.sql.columnar.shuffle.codec zstd
spark.gluten.sql.columnar.shuffle.compressionMode rowvector
spark.gluten.sql.columnar.shuffle.preferSpill true

vanilla spark

gluten

Through more specific logging, we found that this is related to the size of the ;arge buffer eventually used for compression in rowvector mode.
{[471711,41789,11]} means {[uncompressed buffer size, compressed buffer size, uncompressed/compressed]}
1. raw buffer size ~100k

2. raw buffer size ~200k

3. raw buffer size ~450k

Reproduce the shuffle comparison with the listed Gluten compression settings and compare vanilla Spark with Gluten. Start by investigating rowvector-mode compression and the large buffer sizes shown in the logs; done means identifying why Gluten writes more shuffle data and documenting or fixing the cause.

Written by the indexing model from the issue text.