feat(utils): 配置继承、注册表、日志与 Nvidia GPU 守卫

- config: yaml `_base_` 继承 + 深度合并 + 命令行 `--set a.b.c=value` 覆盖，检测循环引用
- registry: 组件按名注册与构建，模型/数据集共用
- device: require_cuda() 守卫，按 CLAUDE.md 约定禁止在无 Nvidia GPU 的设备上训练；
  评估与导出走 resolve_device() 允许 CPU
- labels: multilabel / multiclass 两种图像级标签形式的共用定义
- logging / seed: 控制台+文件日志、随机种子与 cudnn 确定性

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

utils/seed.py

@@ -0,0 +1,33 @@
+"""随机种子与 cudnn 确定性设置。"""
+
+from __future__ import annotations
+
+import os
+import random
+
+
+def seed_everything(seed: int, deterministic: bool = False) -> int:
+    random.seed(seed)
+    os.environ["PYTHONHASHSEED"] = str(seed)
+
+    try:
+        import numpy as np
+
+        np.random.seed(seed)
+    except ImportError:  # pragma: no cover
+        pass
+
+    try:
+        import torch
+
+        torch.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+        if deterministic:
+            torch.backends.cudnn.deterministic = True
+            torch.backends.cudnn.benchmark = False
+        else:
+            torch.backends.cudnn.benchmark = True
+    except ImportError:  # pragma: no cover
+        pass
+
+    return seed