No dashboard shows firing alerts — add an Alertmanager panel

[Gerrrt]

What and why

The stack has 40 alert rules — 32 Prometheus and 8 Loki — and one
Alertmanager routing tree, and no dashboard shows any of them. None of the
five provisioned dashboards references the Alertmanager datasource:

$ grep -l '"alertmanager"' stacks/observability/grafana/dashboards/*.json
(no matches)

The datasource is provisioned and working. Verified through the Grafana proxy:

{ "cluster": { "status": "ready" }, "versionInfo": { "version": "0.33.1" } }

So the only ways to see what is firing today are the Alertmanager web UI or
curl localhost:9090/api/v1/alerts. Neither is where anyone looks. Right now
that hides exactly one warning — IloBatteryCondition, the ProLiant's failed
system battery — which has been firing for over a day without appearing on any
screen anyone opens.

This matters more than it looks, because notification delivery is fragile here
by design: the Alertmanager webhook was pointed at ntfy.example.invalid for
the entire life of the stack, so no alert had ever been delivered until it
was fixed in #31. A dashboard panel is the fallback that does not depend on a
receiver being correctly configured.

Proposal

Add an alert list panel backed by the Alertmanager datasource
(uid: alertmanager). Two reasonable shapes, no strong preference:

• A panel on each existing dashboard, filtered to that dashboard's concern —
  severity/category labels already exist on the rules, and network-snmp
  and ups-power map cleanly onto them.
• One "Alerts" dashboard listing everything firing, grouped by severity,
  with the other five linking to it.

Worth including regardless:

• firing alerts grouped by severity, most severe first;
• how long each has been firing — a warning that has been up for a day reads
  very differently from one that started a minute ago;
• silences, so a silenced alert is visibly silenced rather than absent. An alert
  that vanishes when silenced is how a silence outlives the reason for it.

Notes

• The Alertmanager datasource plugin is frontend-only — its plugin.json
  has no "backend": true, unlike prometheus. So
  /api/datasources/uid/alertmanager/health returns plugin unavailable, and
  that is expected rather than a fault. Do not chase it; the datasource works.
• scripts/check_dashboards.py runs in CI and will hold any new panel to the
  24-column grid with no overlaps, a resolvable datasource UID, and a target on
  every panel. Adding a panel to an existing dashboard means re-flowing the
  gridPos of everything below it.
• Dashboards are provisioned from files with allowUiUpdates: false, so this is
  an edit to the JSON in stacks/observability/grafana/dashboards/, not a
  change made in the UI. docs/roadmap.md already notes that the export loop is
  manual and therefore skipped under pressure.
• Grafana now serves TLS (feat(observability): serve Grafana over TLS with the lab CA #35), which changes nothing here but is worth knowing
  when testing against the API.