The Linear Representation Hypothesis and the Geometry of Large Language Models