Benchmarking Multimodal Large Language Models for Face Recognition