Add queue observability for liaison internal pipelines

[hanahmily]

Problem

Liaison queue behavior is not easy to observe end-to-end during chunked sync pressure and failures.

Current context

• Queue-sub already exposes chunk-ordering/error metrics in banyand/queue/sub/server.go.
• Liaison wires sub.NewServerWithPorts(..., "liaison-server", ...) in pkg/cmdsetup/liaison.go.
• Existing dashboards aggregate some queue errors (banyandb_queue_sub_total_msg_sent_err) in docs/operation/observability.md and docs/operation/grafana-cluster.json.

Proposal

• Add/extend liaison-focused queue metrics (depth, retries, per-topic throughput/latency, failed-part counters).
• Add dashboard panels and alert suggestions for liaison queue health.
• Document metric meanings and troubleshooting paths.

Acceptance criteria

• New metrics are exported and documented.
• Dashboards include liaison queue saturation/failure visibility.
• Integration/e2e validation demonstrates metric changes under injected failures.