Properly shut down distributed process group

This avoids pending NCCL operations being lost.

pytorch-ddp-example/main.py

@@ -289,6 +289,8 @@ def main():
     print0('Final test loss:', test_loss)
     save0(model, 'model-final.pt')
 
+    torch.distributed.destroy_process_group()
+
 
 if __name__ == '__main__':
     main()

pytorch-fsdp-example/main.py

@@ -341,6 +341,8 @@ def main():
     print0('Final test loss:', test_loss)
     save_model(model, 'model-final')
 
+    torch.distributed.destroy_process_group()
+
 
 if __name__ == '__main__':
     main()