VRDU: A Benchmark for Visually-rich Document Understanding