On the Branching Bias of Syntax Extracted from Pre-trained Language Models