Towards Optimal Head-to-head Autonomous Racing with Curriculum Reinforcement Learning