use iter to compare with gwd_start_iter

Summary:
X-link: facebookresearch/FBGEMM#292

`self.step` is used for logging purposes only and is not properly checkpointed/reloaded. Instead, `iter` should be used as it's reloaded from external https://fburl.com/code/48caflz4

Reviewed By: spcyppt

Differential Revision: D63616621

fbshipit-source-id: fb7aacf0ad57088595b8b0d39bee3076b308d394

fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py

@@ -1910,16 +1910,17 @@ def forward(  # noqa: C901
                     ),
                 )
             elif self._used_rowwise_adagrad_with_global_weight_decay:
+                iter_ = int(self.iter.item())
                 apply_global_weight_decay = (
-                    self.step >= self.gwd_start_iter and self.training
+                    iter_ >= self.gwd_start_iter and self.training
                 )
                 return self._report_io_size_count(
                     "fwd_output",
                     invokers.lookup_rowwise_adagrad.invoke(
                         common_args,
                         self.optimizer_args,
                         momentum1,
-                        iter=int(self.iter.item()),
+                        iter=iter_,
                         apply_global_weight_decay=apply_global_weight_decay,
                         prev_iter_dev=self.prev_iter_dev,
                         gwd_lower_bound=self.gwd_lower_bound,