RAR-b: Reasoning as Retrieval Benchmark