The Surprising Agreement Between Convex Optimization Theory and Learning-Rate Scheduling for Large Model Training