Approximated Likelihood Ratio: A Forward-Only and Parallel Framework for Boosting Neural Network Training