* add elementwise div

* move mul and div grad functor

* Combine multiple CUDA kernels

* Update the reduce interface call

* add multi-output

* add multi-output div

* add branch judge

* Package branch

* Combine the x and y functions into one