Tokenization is Sensitive to Language Variation