Stop reliability-rate alerts from exhausting ClickHouse memory and tripping unrelated dashboards

Fixes a ClickHouse OOM that was killing unrelated dashboard and analytics queries with MEMORY_LIMIT_EXCEEDED. Reliability-rate flaky-test alert evaluations are now routed through bucketed materialized views (the same fast path flakiness already used), which drops their per-query memory from multi-GiB to sub-GiB. A per-query max_memory_usage cap (default 6 GiB) is also applied on the read path so any future runaway query fails on its own rather than taking out the rest of the server. Users see reliability-rate rolling windows up to 750 runs evaluating normally again, and dashboards stop failing intermittently while those alerts run.