From Configuration-Space Clearance to Feature-Space Margin: Sample Complexity in Learning-Based Collision Detection