XIFBench: Evaluating Large Language Models on Multilingual Instruction Following