Inferring the Scheduling Policies of an Embedded CUDA GPU