Warning for dataset download & extract race condition

See #8707 and pytorch/pytorch#68320.

Author: EIFY

docs/source/datasets.rst

@@ -27,6 +27,15 @@ All the datasets have almost similar API. They all have two common arguments:
 ``transform`` and ``target_transform`` to transform the input and target respectively.
 You can also create your own datasets using the provided :ref:`base classes <base_classes_datasets>`.
 
+.. warning::
+
+ When built-in datasets are created at the given root directory for the first time,
+ the main process has to download and extract the files in serial. This step is not
+ safe to execute in parallel and will break when multiple processes are launched for
+ distributed training. As a workaround, launch a single process to create the dataset
+ at the given root directory first, terminate it once it is done downloading and
+ extracting the files, and then launch the distributed training.
+
 Image classification
 ~~~~~~~~~~~~~~~~~~~~