Query-time aggregations error on NaN/Inf instead of propagating them like Prometheus #857

Problem

aggregate_values in asap-query-engine/src/engines/query_time_aggregation.rs returns Err("query-time aggregations cannot process non-finite values") when any input value is NaN or ±Inf. It also errors when the result is non-finite. Prometheus instead propagates them: sum of a NaN is NaN, and max/min follow Prometheus's NaN rules.

As a result, one NaN or Inf sample in any group fails the whole query, whereas the same query without a query-time stage returns NaN/Inf. Two paths reach this:

• the legacy planner's nested plans (planned_subquery + query_time_aggregations);
• MILP exact roll-ups (feat(planner): MILP roll-ups for exact sum, count, min and max #852), which regroup exact sum/count/min/max deployments with a query-time stage.

Fix direction

Make each query-time operator follow Prometheus's NaN/Inf semantics instead of erroring. Then add tests comparing against the same query served without a stage.