fix: gpu memory leaks

services/yolo.py

@@ -63,6 +63,10 @@ class YOLOv8Utils:
         # Normalize to [0, 1] (YOLOv8 expects normalized input)
         frame_normalized = frame_resized / 255.0
 
+        # NOTE: Don't track these tensors - they're short-lived inputs to TensorRT
+        # that get automatically freed by PyTorch after inference completes.
+        # Tracking them would show false "leaks" since we can't track when TensorRT consumes them.
+
         return frame_normalized
 
     @staticmethod