!9550 Optimize performance of PyNative grad reduce

From: @jojobugfree
Reviewed-by: @kisnwang,@jjfeing
Signed-off-by: @jjfeing

mindspore/nn/wrap/grad_reducer.py

@@ -65,9 +65,7 @@ def _tensors_allreduce(degree, mean, allgather, allreduce, allreduce_filter, gra
         grad = allreduce(grad)
         if mean:
             degree = F.scalar_cast(degree, F.dtype(grad))
-            cast_op = P.Cast()
-            mul_op = P.Mul()
-            grad = mul_op(grad, cast_op(F.scalar_to_array(1.0 / degree), F.dtype(grad)))
+            grad = F.tensor_mul(grad, F.cast(F.scalar_to_array(1.0 / degree), F.dtype(grad)))
         return grad
     return grad