remove deberta model completely

add Bi-LSTM model weights and vocab

fix: bundle deberta tokenizer files to bypass Hub parsing errors

feat: initial clean project setup (root-level)