feat(observability): attribute dashboard panels per cluster and server

[gabriele-wolfox]

The Grafana dashboard was only validated against a single server and cluster, where sums and maxima are unambiguous. With several servers and clusters it folded independent entities into single misleading values and could not tell them apart.

Add a cluster_name attribute to the plugin backup metrics (klio.plugin.backup.*) so backup panels can be attributed per cluster even when several clusters share a namespace.

Rework the dashboard so every aggregation groups by its identifying label and nothing is folded silently:

• Identify servers by the OpenTelemetry service.name instead of the pod host name, which collides when two servers share a name in different namespaces.
• Scope server panels by service.name and cluster_name rather than the namespace, so a cluster backed up cross-namespace is attributed to the right server.
• Group every panel by cluster, server and tier, and identify each series in its legend, so per-tier backup and WAL state is never folded (tier-1 and tier-2 have independent relay progress and retention).
• Render the PostgreSQL timelines as stepped time series rather than a single current value, so a promotion or failover is visible as the step where the line jumps.
• Derive a cluster label from the Kopia snapshot source so the base snapshot panels are per cluster too.

Update the metric catalog descriptions and the OpenTelemetry and Grafana documentation, and regenerate the committed dashboard JSON.

Assisted-by: Claude

Closes #149