Describir: Scaling Parallel Full-Graph GNN Training to Thousands of GPUs