Persistent Topological Features in Large Language Models