Lantern: predict weights per URL

After some local improvements (#5482 is *amazing*), the optimistic/pessimistic estimates themselves are quite accurate and the next best improvement might be better blending/choosing between them rather than improving them in isolation.

It never made complete sense to blend two times based on graphs and arrive at a time that has nothing to do with the graph. I might prefer trying to predict which estimate is better optimistic/pessimistic and just go with it. If we could do this prediction with 100% accuracy, we could reduce error ~2-6% across the board.