[Dist][Optim] Change op order in SparseAdagrad to be numerically closer to PyTorch (#4253)

Signed-off-by: Serge Panev <spanev@nvidia.com>
Co-authored-by: Mufei Li <mufeili1996@gmail.com>

python/dgl/distributed/optim/pytorch/sparse_optim.py

@@ -255,7 +255,7 @@ class SparseAdagrad(DistSparseGradOptimizer):
             update_event.record()
 
         # update emb
-        std_values = grad_state.add_(eps).sqrt_()
+        std_values = grad_state.sqrt_().add_(eps)
         tmp = clr * grad_values / std_values
         tmp_dst = tmp.to(state_dev, non_blocking=True)