Clone, not copy, the RGB slice of a tensor in tensor_convert_rgb

tensor_convert_rgb is typed for NHWC torch tensors, but with prefer_copy it called image.copy(), which torch.Tensor does not have, so every RGBA image raised AttributeError. Use clone() for a tensor and keep copy() for a numpy array.

Fixes #1194, fixes #1178.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

modules/impact/utils.py

@@ -54,7 +54,8 @@ def tensor_convert_rgb(image, prefer_copy=True):
     if n_channel == 4:
         image = image[..., :3]
         if prefer_copy:
-            image = image.copy()
+            # torch tensors have clone(), not copy(); numpy arrays have copy().
+            image = image.clone() if torch.is_tensor(image) else image.copy()
         return image
 
     if n_channel == 1: