* "derived cudnnDevice context"

* "leave remove cudnn handle from CUDADeviceContext"

* "fix math function error"

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`

* Add HasCUDNN to detect if CUDNN is installed or not

* Fix CI

* change GPU memory allocating policy

* fix potential overflow bug

* Fix Proformance problem of enforce

* Fix missing `;` in code

* Fix CI

* fix compile on cudnn7

* update

* update

* make silent

* conv cudnn 3d

* update test case

* update

* update

* follow comments and remove groups from helper

* update

* refine

* update

* follow comments2

* update

* fix compile

* make enforce a target and dependent on nccl when gpu is enabled

* add some more dependency

* Support int64 for sum op

* Refine code

* "fix accuracy kernel bug"

* "relauch ci"

It caused by a bug of std::call_once described in https://stackoverflow.com/questions/41717579/stdcall-once-hangs-on-second-call-after-callable-threw-on-first-call. It is likely caused by a deeper bug of pthread_once, which is discussed in https://patchwork.ozlabs.org/patch/482350/

CompareOp can run on CPU even other operators are running on GPU, since
opeatations like comparing control flags should be performed only on CPU

* remove unused code

* fix cmake file

* fix build error

* init cudnn batch norm op

* rename batch_norm_cudnn_op.cc batch_norm_op.cu

* correct name style

* add ExtractNCWHD, simplify code

* fix ExtractNCWHD

* use CUDNN_ENFORCE instead of PADDLE_ENFORCE

* "add nccl enforce"

* Dev

* Update comment

* Add nccl test

* Follow comments

* initial matmul operator

Similar to np.matmul, but also has transpose_X and transpose_Y flags,
and only supports tensors from rank 1 to 3 inclusive.

For GPU, uses cublas?gemmStridedBatched. For CPU, uses
cblas_?gemm_batch if available via MKL; otherwise a simple serial
implementation that loops over the batch dimension is employed for now.