Predicting Music Valence and Arousal via Textual and Multimodal In-Context Learning in Large Language Models