Can LLMs Simulate Personas with Reversed Performance? A Benchmark for Counterfactual Instruction Following