Neural Quasiprobabilistic Likelihood Ratio Estimation with Negatively Weighted Data