Democratic or Authoritarian? Probing a New Dimension of Political Biases in Large Language Models