From Where Words Come: Efficient Regularization of Code Tokenizers Through Source Attribution