A GPU-specialized Inference Parameter Server for Large-Scale Deep Recommendation Models