PHOENIX-7962 Fix ReplicationLogGroup mode initialization and handle failover deadlock

[tkhurana]

Summary

A ReplicationLogGroup is a writer and should exist only where the local cluster is active for the HA group. Previously it was created unconditionally — lazily on first mutation and eagerly at RS startup — which let non-active clusters start writers and re-ship (split-brain risk), and let a cutover writer deadlock standby failover by continuously rotating files into the peer's shard directory.

Changes

• Role-gated creation — ReplicationLogGroup.get(...) now returns Optional and creates a group only when the local role is active. A non-active role returns empty and is not cached, so the next call after promotion re-checks. Live write path (preBatchMutate) throws on empty (split-brain guard); replay path (preWALRestore) skips re-ship on empty.
• Skip eager init on non-active clusters — PhoenixRegionServerEndpoint only eager-inits where the local role is active; others init lazily on promotion.
• Failover deadlock fix — a new failoverPending flag (set on entry to the ACTIVE_IN_SYNC_TO_STANDBY cutover gate, cleared on abort) suspends log rotation and turns a SYNC-write failure into an RS abort instead of the illegal cutover→SAF transition. init() seeds the flag on restart-in-cutover.
• Consolidated state listeners — all LOCAL state subscriptions moved into subscribeToStateChanges(), each paired with its unsubscribe so close() leaks no watchers; handles cutover suspend, STANDBY/DEGRADED_STANDBY close, and SYNC ↔ SYNC_AND_FORWARD flips.
• Constructor overload — init() reuses the record the role gate already read.

Tests

9 new unit tests in ReplicationLogGroupTest plus 2 integration tests (testWriterClosesOnDemotionToStandby, testCutoverSuspendsAndResumesRotation).

Note

Behavioral change: a sync-path write to a non-active cluster now throws rather than silently creating a writer — the intended split-brain guard.