Linear Representations of Sentiment in Large Language Models