Reliability-rate alert evaluations no longer exhaust ClickHouse memory

Reliability-rate rolling evaluations now use smaller, pre-sorted bucket materialized views rather than scanning the full 1000-run aggregate, cutting peak memory usage from about 4-5 GiB to sub-GiB. This removes the concurrency spike that was saturating ClickHouse and causing intermittent MEMORY_LIMIT_EXCEEDED failures for dashboards and analytics. Reads also have a per-query memory cap as defense in depth.