Efficient End-to-End Visual Document Understanding with Rationale Distillation