avg calculation with Parquet file in vector runtime seems more "off"

[philrz]

tl;dr

This is a follow-on from #5516. Repeating the avg calculation using the test data from that issue, consider the results in the table below. Treating the result 1058.5234017218875 from DuckDB as a baseline, with the different super formats and runtime options we see:

Format	Runtime	Result	Delta
BSUP	sequential	1058.523401720017	0.0000000018705
CSUP	sequential	1058.523401720017	0.0000000018705
Parquet	sequential	1058.523401720017	0.0000000018705
CSUP	vector	1058.5234017218877	0.0000000000002
Parquet	vector	1058.5297770606794	0.0063753387919

Details

Repro is with super commit 883ffd2.

The the runs that generated the results in the table above are in #5516 (comment) and the test data is linked from that issue.

Users are accustomed to seeing small differences in precision with floating point math, so I expect this might all be explained by something about parallel operations. Indeed, we've seen some non-deterministic floating point calculations with other database systems, for instance. However, seeing the delta as soon as the 4th digit with vector Parquet is surprising compared to out at 9+ digits like the others, so I figured I'd surface this in case it's worthy of closer scrutiny.