* add double grad compute for batch norm,test=develop

* fix unittest, test=develop

* remove unuse tensor,test=develop

* add format,test=develop

* update, test=develop