BeyondWeb: Lessons from Scaling Synthetic Data for Trillion-Scale Pretraining