LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification