Multilingual Multimodal Pre-training for Zero-Shot Cross-Lingual Transfer of Vision-Language Models