Distribution-Dependent Sample Complexity of Large Margin Learning