Taxonomy-Adaptive Moderation Model with Robust Guardrails for Large Language Models