Analogical Reasoning Inside Large Language Models: Concept Vectors and the Limits of Abstraction