Add an interface to set the number of threads for math function, and set the default value to 1 for inference.

stride should be int64_t, not int

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`

* "make global tensor function independently"

* "replace functor"

* "fix inline template error"

* "fix tensor array with CopyFrom"

* "fix other case use CopyFrom"

* "move the op interface hardly"

* "fix operators"

* "fix typo"

* "delete dynamic recurrent rnn and fix gru_unit in debugmode"

* "fix unique_ptr copy"

* "fix cuda copy"

* "fix namespace error"

* "removed nccl python test"

* "fix include error"

* "fix typo"

* fix copy util test

* 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.

* Also use `const DeviceContext&` all the time, to prevent `const_cast`

Fix #4169
Fix #3468
Fix #3475