The Wording Effect: Quantifying Two-Way Drift in LLM Benchmark Performance