fix(detect): 可视化推理自己按批切分，避免 CUDA OOM

model.predict 不理会传入的 batch 参数，把整个图片列表当成一个 batch，8969 张图在
16 GiB 显卡上直接 OOM（单次要分配 10.27 GiB）。改成自己按 --batch 切片调用。

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

scripts/visualize_detections.py

@@ -159,12 +159,19 @@ def main() -> int:
     rows = []
     processed = 0
 
-    for split, images in splits.items():
+    def predict_in_batches(images):
+        """自己按批切分：ultralytics 的 predict 不理会 batch 参数，
+        一次性把整个列表当成一个 batch，上万张图必然 OOM。"""
+        for start in range(0, len(images), args.batch):
+            chunk = images[start:start + args.batch]
             results = model.predict(
-            source=[str(p) for p in images], stream=True, imgsz=args.imgsz, conf=args.conf,
+                source=[str(p) for p in chunk], imgsz=args.imgsz, conf=args.conf,
-            device=device_arg, batch=args.batch, verbose=False,
+                device=device_arg, verbose=False,
             )
-        for image_path, result in zip(images, results, strict=True):
+            yield from zip(chunk, results, strict=True)
+
+    for split, images in splits.items():
+        for image_path, result in predict_in_batches(images):
             predictions = [
                 Box(int(cls), *(float(v) for v in xyxy), float(score))
                 for xyxy, cls, score in zip(