Benchmarking Long-context Document Understanding with Visualizations