test MRG random stream with device=cuda and target=cpu, shared variable + computation in a real model stay on CPU. #6382