Regex-based tokenizer that splits text into tokens with byte offsets.

Whitespace tokens are preserved for continuous offset mapping. No text normalization is applied.