feat: 训练与评估入口及 yaml 配置

- train.py: 先过 require_cuda() 守卫再建模型/数据，支持 --resume 与 --set 覆盖
- eval.py: 允许 CPU，输出定位与分类指标并可写 json
- configs: model/dataset/train/eval 四类，通过 `_base_` 组合；
  classes 既可直接写列表，也可指向 YOLO data.yaml 由框架读 names

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

configs/model/resnet18.yaml

@@ -0,0 +1,15 @@
+# torchvision resnet18：stride 32 + 3 次上采样 => 输出 stride 4。
+model:
+  name: defect_net
+  backbone:
+    name: torchvision
+    arch: resnet18
+    pretrained: true
+    in_channels: 3
+  neck:
+    name: upsample
+    channels: 64
+    num_upsamples: 3
+  head:
+    hidden_channels: 64
+  with_classification: true

configs/model/tiny_conv.yaml

@@ -0,0 +1,18 @@
+# 自研轻量骨架：backbone stride 16 + 2 次上采样 => 输出 stride 4。
+model:
+  name: defect_net
+  backbone:
+    name: tiny_conv
+    in_channels: 3
+    widths: [32, 64, 128, 256]
+    depths: [1, 2, 2, 2]
+    block: dw          # dw=深度可分离(更快) / res=残差(更强)
+  neck:
+    name: upsample
+    channels: 64
+    num_upsamples: 2
+  head:
+    hidden_channels: 64
+  with_classification: true
+  classification:
+    dropout: 0.1