Knowledge Overshadowing Causes Amalgamated Hallucination in Large Language Models