One Forward is Enough for Neural Network Training via Likelihood Ratio Method