feat: emit ReplicaCaughtUp event when streaming replica lag drops to zero

## Problem

When a streaming replica recovers from replication lag — after a network partition, pod restart, or heavy write burst — there is currently no Kubernetes event or signal emitted. Operators must continuously poll `pg_stat_replication` or watch metrics dashboards to know when a replica has fully caught up to the primary. In environments with many clusters and replicas, this creates an observability gap that makes incident response slower and harder to automate.

## Proposed Solution

Emit a `Normal/ReplicaCaughtUp` Kubernetes event on the `Cluster` object the moment a streaming replica's `replay_lag` transitions from non-zero to zero. The event fires exactly once per transition per replica, giving operators a precise, single-fire signal for replication recovery.

This integrates naturally with:
- `kubectl get events` / `kubectl describe cluster`
- Alertmanager and any event-driven alerting pipeline
- Kubernetes event exporters

## Example event output

```
Normal ReplicaCaughtUp cluster/pg-cluster Replica pg-cluster-2 has caught up to the primary
```

## Implementation approach

- Track per-replica lag state in an in-memory map on `ClusterReconciler` (never written to the public CRD status, so the API schema stays clean)
- Call `detectAndEmitReplicaCaughtUpEvents` from `updateClusterStatusThatRequiresInstancesState` on every reconcile loop
- Protect the tracker with a `sync.Mutex` for safe concurrent reconciliation of multiple clusters
- Zero API changes — `InstanceReportedState` and all public types are untouched

A PR with implementation and unit tests is ready: #11291

Inizia esaminando PR #11291 e l’entry point updateClusterStatusThatRequiresInstancesState, quindi segui il modo in cui detectAndEmitReplicaCaughtUpEvents tiene traccia del ritardo delle repliche ed emette eventi Kubernetes. Usa gli unit test inclusi nell’implementazione per verificare che venga emesso un singolo evento quando replay_lag passa da un valore diverso da zero a zero per ogni replica.

Scritto dal modello di indicizzazione a partire dal testo della issue.