Created by: wangkuiyi

We used to support only CPU and CUDA GPU. In that context, PADDLE_ONLY_CPU == ! PADDLE_WITH_CUDA.

However, if we are going to support more devices, including FPGA and non-CUDA GPUs, the theorem PADDLE_ONLY_CPU == ! PADDLE_WITH_CUDA doesn't stand. Instead, we need PADDLE_WITH_CUDA.