fix(sync): 限流退避持久化到 DB，调度器与手动同步均尊重退避期

db.py: 修复迁移非幂等——gunicorn 双 worker 并发 init_db 时后到者遇 Duplicate column 会让整个服务起不来，改为吞掉 duplicate column 错误。garmin.py/scheduler.py: 退避期写入 sync_status.rate_limited_until，自动同步与手动同步在退避期内跳过，不再反复撞击 Garmin 把限流窗口越撞越深。

backend/services/scheduler.py

@@ -144,6 +144,16 @@ def sync_all_accounts(days=None, respect_schedule=False):
                     results.append({"user": uid, "status": "skipped",
                                     "reason": "not due"})
                     continue
+            # Never poke Garmin while it is rate-limiting us — that is exactly
+            # what keeps the limit alive. Respect the persisted cooldown and sit
+            # this tick out.
+            blocked = garmin_svc.rate_limited_until(uid)
+            if blocked and blocked > _now():
+                results.append({
+                    "user": uid, "status": "skipped", "reason": "rate-limited",
+                    "retryAfterSeconds": int((blocked - _now()).total_seconds()),
+                })
+                continue
             out = garmin_svc.sync_data(uid, {}, days=days)
             results.append({"user": uid, "status": out.get("status"),
                             "records": out.get("recordsSynced")})