Specializing General-purpose LLM Embeddings for Implicit Hate Speech Detection across Datasets