RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content