[PyTorch] Fix bug with clearing op outputs during backward (#2008)

Fix merge conflict bug with clearing op outputs
Signed-off-by: Tim Moon <tmoon@nvidia.com>

transformer_engine/pytorch/ops/fuser.py

@@ -199,7 +199,7 @@ class _OperationFuserAutogradFunction(torch.autograd.Function):
 
         # Mark output tensors as not deletable in backward
         for tensor in [x] + extra_outputs_flat:
-            tensor.do_not_clear = True
+            tensor._do_not_clear = True
 
         x.requires_grad_(fuser.first_op_requiring_backward < fuser._num_basic_ops)