Scaling Agent Learning via Experience Synthesis