Multimodal Large Language Models for Low-resource Languages: Cross-lingual Reasoning, Retrieval-augmented Generation, and Hallucination-aware Evaluation