Game-Theoretic Regularized Self-Play Alignment of Large Language Models