Created by: jacquesqiao

to support running in a multi device environment, for example, the formal op is running in a GPU kernel, and current op only support CPU kernel, we need some functions to transform different data type. These functions should be implement and registered to the framework, such as we have a global map

map<KernelTypePair, DataTransformationFN> g_data_transformation_fn_;

and will be get a proper tranform function when call operator::run()