ContraDoc: Understanding Self-Contradictions in Documents with Large Language Models