perf(request-logs): serve listing totals from the demand rollup plus raw tail (#1615)

The listing total was an exact COUNT(*) over the filtered set — a full
pass over request_logs (4.4s warm, 9-19s under contention at 4.5M rows)
re-run every 30s when the count cache expires at the dashboard's poll
interval. The demand quarter rollup carries status as a dimension, so
every listing filter except free-text search and error-code splits maps
onto it: count the folded window as one SUM(request_count) bounded by the
hourly watermark (new sum_demand_window primitive) and only the un-folded
complement from raw. Totals stay exactly equal to the raw count for every
watermark position; no watermark degrades to the legacy count.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>