Understanding the Training Speedup from Sampling with Approximate Losses