Tight Sample Complexity of Large-Margin Learning