Avoid inplace modification of target boxes in detection references (#5289)

@datumbox
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

references/detection/coco_utils.py

@@ -156,7 +156,7 @@ def convert_to_coco_api(ds):
         img_dict["height"] = img.shape[-2]
         img_dict["width"] = img.shape[-1]
         dataset["images"].append(img_dict)
-        bboxes = targets["boxes"]
+        bboxes = targets["boxes"].clone()
         bboxes[:, 2:] -= bboxes[:, :2]
         bboxes = bboxes.tolist()
         labels = targets["labels"].tolist()