Revisiting Inferential Benchmarks for Knowledge Graph Completion