Challenging Multilingual LLMs: A New Taxonomy and Benchmark for Unraveling Hallucination in Translation