Scheduling the Learning Rate via Hypergradients: New Insights and a New Algorithm