Towards Quantized Model Parallelism for Graph-Augmented MLPs Based on Gradient-Free ADMM Framework