Describe the feature and motivation

I used Yolov8's onnx model for target detection through C++, and the detection results running through the CPU were correct. When calling net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV) and
net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU) turns on CUDA DNN detection, the data returned by the net.forward() function, the class_id and confidences data are consistent with the CPU operation, but all box data are null.
CPU result:

GPU result：

The OPENCV version I used is 4.8.1, the CUDA version is 11.8, the CUDNN version is 8.9.6, the graphics card model is NVIDIA GeForce GTX 1660 Ti, and the driver version is 546.17.

In addition, if the model is called directly in Python, the results of running on CPU or GPU will be correct. The same OpenCV, CUDA and DNN versions are used when running in Python and C++.

Additional context

No response