Scaling Law for Recommendation Models: Towards General-purpose User Representations