Created by: wzzju

PR types

Others

PR changes

Others

Describe

• Remove sorted_sum_gradient_ form BasicEngine and PartialGradTask.
• Use Tensor instead of Variable in the doc of paddle.grad.