Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL