ARB: A Matched Authorship-Rewriting Benchmark Dataset for AI-Text Detector Evaluation