A Modular Taxonomy for Hate Speech Definitions and Its Impact on Zero-Shot LLM Classification Performance