Rethinking Relation Extraction: Beyond Shortcuts to Generalization with a Debiased Benchmark