feat(engine): 训练评估循环、CenterNet 损失与指标

- trainer: AMP、梯度裁剪、cosine/step 调度、last/best 检查点，按 monitor 指标挑选 best
- losses: penalty-reduced focal loss（heatmap）+ 掩码 L1（wh/offset）+ 分类损失
  （多类 BCE / 单类 CE，随 label_mode 切换）
- metrics: 定位 IoU 匹配下的 P/R/F1 与 mean IoU；分类多标签 micro-F1 与单标签 accuracy，
  两种模式都输出 cls_f1，故 train.monitor 无需跟着改
- optim: BatchNorm 与 bias 不做 weight decay
- visualizer: SwanLab 记录损失/学习率/验证指标/显存峰值，未安装或离线时降级为警告

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

engine/checkpoint.py

@@ -0,0 +1,46 @@
+"""检查点存取。"""
+
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any
+
+import torch
+from torch import nn
+
+
+def save_checkpoint(
+    path: str | Path,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer | None = None,
+    epoch: int = 0,
+    best_metric: float | None = None,
+    config: dict | None = None,
+) -> Path:
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    torch.save(
+        {
+            "model": model.state_dict(),
+            "optimizer": optimizer.state_dict() if optimizer else None,
+            "epoch": epoch,
+            "best_metric": best_metric,
+            "config": config,
+        },
+        path,
+    )
+    return path
+
+
+def load_checkpoint(
+    path: str | Path,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer | None = None,
+    map_location: Any = "cpu",
+    strict: bool = True,
+) -> dict:
+    state = torch.load(Path(path), map_location=map_location, weights_only=False)
+    model.load_state_dict(state["model"], strict=strict)
+    if optimizer is not None and state.get("optimizer"):
+        optimizer.load_state_dict(state["optimizer"])
+    return state