Improving the Scaling Laws of Synthetic Data with Deliberate Practice