fix: parse SHOW REPLICAS data_info correctly; add replication health + drift metrics

[nitinstp23]

Summary

While debugging intermittent empty query results in a production cluster, we found a replica that had been registered and heartbeating for 104+ days while
replicating zero data (SHOW REPLICAS → data_info: {}). The operator never surfaced this: the parser read the wrong column, and both the ReplicationHealthy
condition and existing metrics were based on registration count, not actual data streaming.

Changes

Parser fix (FIXES-REQUIRED.md Issue 1) — internal/memgraph/client.go

• Parse the current Memgraph format (name | socket_address | sync_mode | system_info | data_info), with per-row fallback for the legacy pre-v2.21 format
• Parse data_info into per-database {behind, status, ts}; empty data_info ({}/Null) classifies as invalid — it means data streaming never engaged, not a parse
  artifact
• ReplicaInfo gains SocketAddress/SyncMode/DataInfo/Behind/Status/Timestamp plus IsHealthy() / DataInfoPresent()

New replication metrics — internal/controller/metrics.go

• Per-replica: memgraph_replica_healthy, memgraph_replica_status, memgraph_replica_behind_count, memgraph_replica_last_confirmed_timestamp_seconds,
  memgraph_replica_data_info_present
• Rollups: memgraph_cluster_replicas_total, memgraph_cluster_replicas_healthy_total; memgraph_replication_healthy now means "all registered replicas streaming"
• Drift: memgraph_replication_vertex_drift / memgraph_replication_edge_drift (main minus replica, from existing SHOW STORAGE INFO collection)
• Stale series cleanup via DeletePartialMatch on replica removal and cluster deletion

Condition correctness — ReplicationHealthy now requires every registered replica to be streaming (healthy >= registered), not just registered >= expected; message
reports both counts

Breaking / behavioral notes

• Clusters with registered-but-not-streaming replicas will now (correctly) report ReplicationHealthy=False and emit unhealthy-replica warnings — this is the
  intended detection of the production failure mode
• MetricsRecorder.RecordReplicationHealth replaced by RecordReplicationLag (healthy signal moved to RecordReplicaSet)

Testing

• make lint — 0 issues; make test — all packages pass (incl. envtest suite)
• Parser table tests cover: current format healthy/empty/Null/multi-database/negative-behind, legacy format, malformed rows
• Metrics tests assert values, label sets, and series cleanup via prometheus/testutil

Rollout

Version bumped to 0.2.0. Helm chart 0.1.4 (appVersion bump only) and infra canary (single cluster via per-cluster values override; collector scrape config added
for the one cluster missing it) go in separate PRs.