Support multi-partition groupby standard deviation aggregations

[Matt711]

Description

Supports streaming groupby std and var aggregations.

• Contributes to [Story] Eliminate fallbacks (single partition) for PDS-DS in the rapidsmpf streaming executor #21749

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wence-]

This is mathematically correct but numerically tremendously unstable. We really want the grouped aggregation machinery to expose the "intermediate" value before apply finalisation. The one-pass approach uses welford's algorithm iirc. If we xan get that out then we could use it as an input to merge two grouped objects and then continue

[Matt711]

This is mathematically correct but numerically tremendously unstable. We really want the grouped aggregation machinery to expose the "intermediate" value before apply finalisation. The one-pass approach uses welford's algorithm iirc. If we xan get that out then we could use it as an input to merge two grouped objects and then continue

Yeah I even had to switch around some of the OOOs to get the test to pass (trying to avoid small numerical precision differences).

The alternative you're suggesting is interesting. Reading variance algorithims...

[Matt711]

Interesting, so we already have M2 and MERGE_M2 aggregations in libcudf. I wonder if we can reuse them here.

#21962 computes M2 on each partition and MERGE_M2 on the partial results.