M3-SLU: Evaluating Speaker-Attributed Reasoning in Multimodal Large Language Models