Language as a Label: Zero-Shot Multimodal Classification of Everyday Postures under Data Scarcity