Fixed double-counting in rolling flaky and reliability alerts

Rolling flaky-test and reliability alerts in Tuist now count each test run only once when evaluating thresholds. Previously, the rolling-window evaluation could double or triple-count a single run when ClickHouse re-inserted it after the flaky detector marked it, which inflated flakiness rates and deflated reliability rates and could auto-skip or mute tests that were actually healthy. The fix is read-time (no schema changes or backfill), applies to flakiness_rate, flaky_run_count, and reliability_rate rolling monitors, and now also reads a slightly larger data bucket than the alert window to keep de-duplication headroom.