Bayesian Neural Scaling Law Extrapolation with Prior-Data Fitted Networks