Representation noising effectively prevents harmful fine-tuning on LLMs