Evaluation data contamination in LLMs: how do we measure it and (when) does it matter?