Wukong: Towards a Scaling Law for Large-Scale Recommendation