Navigating the Ocean of Biases: Political Bias Attribution in Language Models via Causal Structures