* Unify the gpu implementation of stack and unstack to reuse the optimization.

* Optimize the cuda implementation of unstack.

* Use GpuMemcpyAsync instead of memory::Copy.

* Fix error of calculating the index.

* Use FastDivMod to further imporve the performance of unstack.

* Implement a common PointerArray.

* Polish codes.

* Add including of header file.

* Add the branch of kFix8.

* Fix compiling error.

* Add alignas hint to fix the performance drop.

* Optimize the H2D copy in stack_grad.

* Rename the macro.

* Fix align hint for different compilers.

* Polish the define of PADDLE_ALIGN.

* Fix compiling error.

* Remove the align hint on windows.

* add stack grad kernel optimization

* add basic optimization kernel for stack_grad_kernel

* optimization of stack_grad_kernel for last dim stack and change code format with pre-commit

* revert concat and change concat to stack

* let stack kernel support int8, uint8 and bool type

* first commit.

* refine performance with fast_divmod

* refine performance with fast_divmod

* refine cmake of framework

* add deps for dense tensor

* fix deps

* remove alloc(ctx)

* add depends on mkldnn

resort .cu headers, set clang-format not sort include block and consider .cu as main source file (#43633)