[Improvement] Aggregation and partial-update merges build a one-row batch per key #1336

Reading a primary-key table with 2M keys written in 3 overlapping rounds takes about 1.97 s with merge-engine=aggregation (sum) and 1.42 s with partial-update, against 0.44 s for deduplicate (release build, in-memory FileIO, medians of 8 reads). Writes show the same gap since the write buffer uses the same merge. In the profile, malloc, GenericByteArray::slice and Arc<dyn Array> drops are near the top: merge_ordered and merge_unreduced return a one-row RecordBatch per key, built from slice(row, 1) for key columns and a one-row array from each aggregator's result().

Batching those one-row batches before interleave made no difference, so the cost is in building them. One option is letting aggregators append their result to a column builder, so a merged key adds a row to per-column builders instead of creating arrays. That touches every aggregator, so I wanted to check whether this direction makes sense before opening a PR.