Homotopy Relaxation Training Algorithms for Infinite-Width Two-Layer ReLU Neural Networks