Boosting Distributed Full-graph GNN Training with Asynchronous One-bit Communication