!6430 fix sparse loss scale

Merge pull request !6430 from riemann_penn/fix_sparse_loss_scale

mindspore/nn/wrap/loss_scale.py

@@ -48,6 +48,9 @@ grad_overflow = P.FloatStatus()
 def _tensor_grad_overflow(grad):
     return grad_overflow(grad)
 
+@_grad_overflow.register("RowTensor")
+def _tensor_grad_overflow_row_tensor(grad):
+    return grad_overflow(grad.values)
 
 class DynamicLossScaleUpdateCell(Cell):
     r"""