* init elem_max_grad op

* optimize code and reply review comments

* ternary functors

* apply new reduce func

* move functor to .h

* multi-outputs init

* rearrange code

* modifed functors

* optimizer code

* pass nullptr

* revert the last change as seg fault occurs

* optimize code

* remove inplace

* remove comments

* ini commit

* multi-outputs init commit

* optimize code

* remove inplace

* fix the wrong filename

* first commit

* first commit

* remove rest useless headers

* for ci approval

* support vs2019 compilation in windows

* not modify pow_op's original compute logic

* init commit: new elem_mul_grad

* add template speciallization for complex in multiply

* reply review comments

* correct dx and dy computation when T is complex

* reply review comments

* update to new ReduceRunctor

* mul-output broadcast

* call functions

* call functions with comments

* remove comments

* 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

* change 'math' to 'math_kernel'

* fix compile bugs

* merge develop

* fix compile bugs

* move cpu_impl of elementwise kernel to new directory

Add elementwise_fmax and elementwise_fmin operators