The Geometry of Persona: Disentangling Personality from Reasoning in Large Language Models