pub fn load_tokenizer(model_dir: &Path) -> Result<Tokenizer>
Load tokenizer: try tokenizer.json first, fall back to vocab.txt (BERT-style).