* pass stop_gradient for cast op

* improve performance of elementwise_add grad

* use tensor copy async

* dygraph branch

* fix dygraph branch

* add ut