add addmm op

* Add the dynamic load of nvrtc, and support runtime compiling of CUDA kernel using nvrtc.
test=develop

* Call CUDA driver api to launch the kernel compiled by nvrtc.
test=develop

* Disable for mac and windows.
test=develop

* Refine the codes to support manually specified num_threads and workload_per_thread.
test=develop

* Refine the CUDA kernel to support large dims.
test=develop

test=develop

* refine cublase
test=develop

* code refine

* refine cublas

* add GEMME_EX

* add enable_cublas_tensor_op_math doc and add cublasCall
test=develop

* fix CublasCall for cuda version
test=develop

* fix error
test=develop

* fix GEMM_EX to be compatible with gcc 4.8
test=develop

* add GEMM_EX
test=develop

* to compatiable with gcc4.8
test=develop

* flags

* "follow comment"

* fix gemm error for cuda 7.5

* fix version number

* Update source files.

* Update headers

* Update

* Update

* Update

* Update

* Fix a CMake dependency

* change from hgemm to gemmEx

* fix cpplint

* test cpu float16 data transform

* add isnan etc

* small fix

* fix containsNAN test error

* add data_type transform GPU test

* add float16 GPU example

* fix error

* fix GPU test error

* initial commit

* fix error

* small fix

* add more gemm fp16 tests

* fix error

* add utility function

* 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.

* Should always use `dynload::` for cuda function.
* Fix cublas.h without DSO load.