Decoding Memes: Benchmarking Narrative Role Classification across Multilingual and Multimodal Models