Llms, Virtual Users, and Bias: Predicting Any Survey Question Without Human Data