Learning to Order: GNN-Based Deep Reinforcement Learning Policies for Autoregressive Graph Models