The Rarity Blind Spot: A Framework for Evaluating Statistical Reasoning in LLMs