scanColumn aggregates: coalesce N aggregates on one column into a single stream

[philcunliffe]

Surfaced during review of icebird PR #28 (the scanColumn streaming hook), part of the hypaware bounded-query-execution effort.

Observation

tryColumnScanAggregate / scanColumnAggregate calls the source scanColumn(column) hook once per aggregate spec. So a query with multiple aggregates over the same column — e.g. SELECT COUNT(c), MIN(c), MAX(c), SUM(c), AVG(c) FROM t — opens N independent scanColumn streams, re-reading and re-decoding that column N times (N× metadata/IO, costly for remote/parquet sources). Memory stays O(1) per stream, but the IO is multiplied.

This is a memory-for-IO tradeoff that only appears once a source implements scanColumn (icebird now does). It is not a correctness bug and does not affect the memory-bounding goal.

Suggested fix

In tryColumnScanAggregate, group the column-scan aggregate specs by { column, limit, offset } and drive each group from a single scanColumn stream, feeding all aggregates for that column from the one pass.

Refs

• icebird PR Execute cleanup #28 (the hook this concerns)
• hypaware bounded-query-execution: design LLP 0058, decision LLP 0055 (streaming aggregates via scanColumn)