Improving Cross-task Generalization of Unified Table-to-text Models with Compositional Task Configurations