Closest Positive Cluster Loss: Improving the Generalization of Implicit Hate Speech Classifiers Across Social Media Datasets