Look Before you Leap: Estimating LLM Benchmark Scores from Descriptions