修复：P0+P1 生产稳定性和性能优化（6项）

P0 稳定性修复：
- 告警去重字典添加惰性清理机制，防止长时间运行内存溢出
- Redis 连接断开时显式 close() 后再置 None，防止文件描述符泄漏
- 截图消息 ACK 移至成功路径，失败消息留在 pending list 自动重试

P1 性能优化：
- GPU NMS 添加 torch.no_grad() + 显式释放临时张量，减少显存碎片
- 截图存储改为 Redis 原始 bytes，去掉 Base64 编解码开销（兼容旧格式）
- ROI 配置查询 N+1 改为 get_all_bindings() 单次 JOIN 查询

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

core/postprocessor.py

@@ -78,22 +78,24 @@ class NMSProcessor:
         max_output_size: int
     ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
         """GPU 加速 NMS"""
-        boxes_t = torch.from_numpy(boxes).cuda()
-        scores_t = torch.from_numpy(scores).cuda()
-        
-        keep = torch_nms(boxes_t, scores_t, iou_threshold=self.nms_threshold)
-        
-        keep_np = keep.cpu().numpy()
-        
-        if len(keep_np) > max_output_size:
-            top_k = np.argsort(scores[keep_np])[::-1][:max_output_size]
-            keep_np = keep_np[top_k]
-        
-        return (
-            keep_np.astype(np.int32),
-            scores[keep_np],
-            class_ids[keep_np] if class_ids is not None else np.array([])
-        )
+        with torch.no_grad():
+            boxes_t = torch.from_numpy(boxes).cuda()
+            scores_t = torch.from_numpy(scores).cuda()
+
+            keep = torch_nms(boxes_t, scores_t, iou_threshold=self.nms_threshold)
+
+            keep_np = keep.cpu().numpy()
+            del boxes_t, scores_t, keep
+
+            if len(keep_np) > max_output_size:
+                top_k = np.argsort(scores[keep_np])[::-1][:max_output_size]
+                keep_np = keep_np[top_k]
+
+            return (
+                keep_np.astype(np.int32),
+                scores[keep_np],
+                class_ids[keep_np] if class_ids is not None else np.array([])
+            )
     
     def _process_cpu(
         self,