Don't pickle an empty dict in LayerNorm and pt base modules (#2253)

Don't pickle an empty dict in LayerNorm and BasicOperation layers
Signed-off-by: Peter St. John <pstjohn@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

transformer_engine/pytorch/ops/op.py

@@ -595,6 +595,9 @@ class BasicOperation(FusibleOperation, metaclass=abc.ABCMeta):
                 extra[key] = val
             state[mode]["extra_fp8_variables"] = extra
 
+        if not state:
+            return torch.empty(0, dtype=torch.uint8)
+
         # Serialize state into byte tensor
         torch.cuda.synchronize()
         state_serialized = bytearray(pickle.dumps(state))