Query engine: roll-ups over keyed aggregations (aggregated_labels) #846

Follow-up to #839, which covers roll-ups only for single-subpopulation configs (KLL, DDSketch, HLL, Sum, MinMax: grouping_labels = G, one sketch per store key).

Goal

Let the engine serve a query grouped by Q from a keyed aggregation whose aggregated_labels ⊋ Q. Keyed configs use grouping = [] and aggregated = G, with a DeltaSet key aggregation or a self-keyed value: MultipleSum, MultipleMinMax, MultipleIncrease, CountMinSketch, CountMinSketchWithHeap and HydraKLL.

Why it's a different mechanism

The subpopulations live inside one sketch, so a roll-up can't merge store keys. Instead it has to enumerate the G-keys (from the DeltaSet, or the accumulator's get_keys), project them onto Q, and combine the per-key answers:

• MultipleSum / MultipleMinMax: an exact sum or min/max of the per-key values. The MILP already gets this through query_time_aggregations as of Query engine: merge sketches across subpopulations (roll-up to a coarser grouping) #839, so the only open question is whether the engine should do it natively.
• CMS: sum the point estimates. The error grows with the fan-in, so it has to be priced or bounded.
• MultipleIncrease, CMS+heap: not mergeable across groups. The counter-reset merge only works across time, and the top-k heaps drop items.
• HydraKLL: the engine wraps a single-key Hydra. sketchlib's full Hydra framework expands each record into its subpopulation subsets, which would answer subsets natively.

Open

Is this worth doing natively in the engine, beyond what query_time_aggregations already covers for the exact types?