* Refactor TensorAdd func by template and remove gradient_accumulation in eager

* Remove needless target name

* Use overload instead of template