Emit first-party cluster-health metrics and ship a default alert-rule set in the Helm chart

[lllamnyp]

The operator currently exposes only stock controller-runtime metrics — the chart ships a ServiceMonitor, and the Cozystack packaging adds a VMServiceScrape, so the scrape surface exists — but no first-party health metrics and no alerting rules. That combination leaves the most damaging failure mode invisible: as #343 established, a cluster can enter a state the reconciler can never leave (e.g. scaleUp with zero live endpoints while status.clusterID is latched), and the loop retries forever while status.conditions keeps reporting Available=True. Anything watching conditions sees green; the only truthful signal is the reconcile loop itself.

Proposal, in two halves:

1. First-party metrics

Per EtcdCluster, labelled namespace/name (names are a strawman):

• etcd_operator_cluster_members_desired / _members_ready / _members_learners
• etcd_operator_cluster_quorum — 0/1 as currently probed by the operator, not as latched in status
• etcd_operator_cluster_condition{condition=...} — terminal/degraded states (AllMembersLost from fix(controllers): report total member loss instead of spinning #343, BootstrapFailed, …)
• etcd_operator_cluster_last_successful_reconcile_timestamp_seconds — the anti-lying heartbeat: consumers alert on its age instead of trusting conditions

2. A default alert-rule set shipped by the Helm chart

Values-gated (e.g. alerts.enabled), following the pattern cloudnative-pg and piraeus charts already use — the chart ships the rules, the cluster's evaluator just selects them:

• Reconcile-error streak: rate(controller_runtime_reconcile_errors_total{controller=~"etcdcluster.*"}[10m]) > 0 sustained 1h → warning, 6h → critical. Catches every wedged-forever class, including the fix(controllers): report total member loss instead of spinning #343 one, on day one — no new instrumentation needed.
• Terminal/degraded condition present → critical immediately.
• members_ready < members_desired for 15m → critical.
• quorum == 0 for 5m → critical.
• time() - last_successful_reconcile > 1h per cluster → warning.

The upstream chart already ships prometheus-operator kinds (ServiceMonitor), so PrometheusRule fits there; victoria-metrics-operator conversion makes it consumable in Cozystack deployments, or the Cozystack package can mirror it as a VMRule next to its existing VMServiceScrape.

Separable follow-up: the operator holds the client TLS material for the etcd clusters it manages, so it is also the natural place to optionally render a per-cluster scrape plus the standard etcd data-plane rules (leader changes, fsync latency, db size vs quota). The five rules above don't depend on it.