Add MSHADOW_FORCE_STREAM preprocessor define (#186)

This change brings the CMake build closer to parity with the upstream
Makefile, which unconditionally enables the MSHADOW_FORCE_STREAM preprocessor define at
the MXNet level. This is kept within the USE_CUDA condition based on the
assumption that it only impacts GPU training.

cmake/mshadow.cmake

@@ -30,6 +30,7 @@ if(USE_CUDA)
 		message(FATAL_ERROR "-- CUDA is disabled.")
 	endif()
 	add_definitions(-DMSHADOW_USE_CUDA=1)
+	add_definitions(-DMSHADOW_FORCE_STREAM)
 	include_directories(SYSTEM ${CUDA_INCLUDE_DIRS})
     list(APPEND mshadow_LINKER_LIBS ${CUDA_CUDART_LIBRARY}
                               ${CUDA_curand_LIBRARY} ${CUDA_CUBLAS_LIBRARIES})