fix: harden message seq allocation against stale redis replicas

- use sentinel failover client for group message runtime when
  sentinel_master_name/sentinel_addrs are configured, so seq counter
  reads never hit a stale replica after failover
- re-seed missing redis seq counter from DB topic seqid instead of
  trusting the caller in-memory baseline
- on unique-key conflict during async persist, re-allocate seq from DB
  latest, bump redis counter and fix the idempotency record so resends
  with the same client mid return the final seq
- add uk(topic, seqid) unique keys for im_messages in tenant DDL

scripts/IM-租户业务库初始化.sql

@@ -1109,7 +1109,8 @@ CREATE TABLE IF NOT EXISTS `im_messages` (
     `head` json DEFAULT NULL,
     `content` json DEFAULT NULL,
     PRIMARY KEY (`id`),
-    UNIQUE KEY `messages_sender_client_mid` (`from`,`client_msg_id`)
+    UNIQUE KEY `messages_sender_client_mid` (`from`,`client_msg_id`),
+    UNIQUE KEY `uk_im_messages_topic_seq` (`topic`,`seqid`)
     ) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_0900_ai_ci;
 
 CREATE TABLE IF NOT EXISTS `im_monthly_reports` (