Hourly rollup tables for gateway activity and uptime #136

Problem

Gateway metrics are computed from raw history on every refresh. The activity query joins transactions → LN outputs → contract details → inputs across the whole window, and uptime and the trend scan every 5-minute poll snapshot in the window. On a large federation, 90D takes ~4 s.

#135 serves the gateway page from a precomputed in-memory overview, but the background refresh still re-runs these queries. On the shared staging/production VPS, one refresh cycle over 17 federations took ~75 s and raised CPU from ~27% to ~35%. #135 now refreshes 30D/90D less often, which helps, but the cost of each run still grows with history.

Relevant files:

• fmo_server/src/federation/gateways.rs (list_federation_gateways, federation_gateway_uptime_trend, fetch_and_store_gateways)
• fmo_server/src/federation/observer.rs (session processing, where LN outputs/inputs are stored)

Proposed changes

Keep small hourly rollups that are updated incrementally, and answer every window by summing them:

• gateway_activity_hourly (federation_id, gateway_key, hour, fund_count, settle_count, cancel_count, volume_msat): updated as sessions are processed (a fund, settle or cancel attributed to a gateway adds to its current hour).
• gateway_uptime_hourly (federation_id, gateway_id, hour, seen_samples, total_samples): updated by each gateway poll, alongside gateway_poll_snapshots.
• Windows (24h/7d/30d/90d) become SUM over at most 2,160 rows per gateway. The daily trend is the same rows grouped by day.
• Hourly rather than daily buckets, so the rolling 24h window stays accurate to within an hour.
• One-time backfill migration from existing transactions / gateway_poll_snapshots.
• With windows this cheap, the overview could be computed on request and the background refresh dropped.

The same tables would serve future per-gateway detail pages and LNv2 activity (#132).

Acceptance criteria

• Every window answers in milliseconds, and the cost no longer grows with history.
• Rollup results match the current queries on a production snapshot (counts, volume, seen/total samples per gateway and window), checked before switching over.
• The retired-gateway rule (no samples counted 7 days after last_seen) still holds.
• Backfill is resumable and safe to re-run.
• CPU on the VPS stays at or below its level before Gateway page: precomputed overview endpoint, faster windows, stable view #135.