Analyzing Political Bias in LLMs via Target-Oriented Sentiment Classification