BenchMIRT: What are LLM benchmarks actually measuring?