Measuring What Matters: Construct Validity in Large Language Model Benchmarks