fix: monitoring stack reliability — cadvisor restart policy and pgss dashboard resilience

- Add `restart: unless-stopped` to self-cadvisor service (the only
  long-running service missing it, causing it to stay dead after
  crash/reboot)

- Add `or` fallback to all 20 pgss Grafana queries in Dashboard_2
  so panels show data even when monitoring_flask_backend is down.
  Previously, when pgwatch_query_info was stale, the mandatory
  group_left() join returned empty results, making all pgss panels
  blank despite data being present in VictoriaMetrics.

Relates to https://gitlab.com/postgres-ai/postgresai/-/issues/148

https://claude.ai/code/session_01SzJxzZNQjDQphaHyaX3RU7

Author: claude