Semantic Redundancies in Image-Classification Datasets: The 10% You Don't Need