LIFEBench: Evaluating Length Instruction Following in Large Language Models