enforce shuffling before sharding (#5680)

* enforce shuffling before sharding

* revert test changes and add comment

torchvision/prototype/datasets/_builtin/svhn.py

@@ -72,6 +72,6 @@ class SVHN(Dataset):
         dp = resource_dps[0]
         dp = Mapper(dp, self._read_images_and_labels)
         dp = UnBatcher(dp)
-        dp = hint_sharding(dp)
         dp = hint_shuffling(dp)
+        dp = hint_sharding(dp)
         return Mapper(dp, self._prepare_sample)

torchvision/prototype/datasets/_builtin/usps.py

@@ -49,6 +49,6 @@ class USPS(Dataset):
     ) -> IterDataPipe[Dict[str, Any]]:
         dp = Decompressor(resource_dps[0])
         dp = LineReader(dp, decode=True, return_path=False)
-        dp = hint_sharding(dp)
         dp = hint_shuffling(dp)
+        dp = hint_sharding(dp)
         return Mapper(dp, self._prepare_sample)

torchvision/prototype/datasets/_builtin/voc.py

@@ -154,8 +154,8 @@ class VOC(Dataset):
         split_dp = Filter(split_dp, functools.partial(self._is_in_folder, name=self._SPLIT_FOLDER[config.task]))
         split_dp = Filter(split_dp, path_comparator("name", f"{config.split}.txt"))
         split_dp = LineReader(split_dp, decode=True)
-        split_dp = hint_sharding(split_dp)
         split_dp = hint_shuffling(split_dp)
+        split_dp = hint_sharding(split_dp)
 
         dp = split_dp
         for level, data_dp in enumerate((images_dp, anns_dp)):