Spend Less, Fit Better: Budget-Efficient Scaling Law Fitting via Active Experiment Selection — ThinkLLM