Multimodal Contrastive Learning of Urban Space Representations from POI Data