The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence