fix: resolve Bi-LSTM model loading error and adjust frontend input UI to fit single page

remove deberta model completely

update backend to Bi-LSTM model and fix paths

fix: correct inverted label mapping

fix: resolve softmax crash in AI detection by accessing .logits attribute

fix: resolve AI detection failure by loading model directly and keep DeBERTa veracity engine

feat: revert core engine to DeBERTa-v3 and fix accuracy by matching training input (content only)

feat: switch to Bi-LSTM for maximum accuracy and fix training-inference mismatch

feat: implement dual-input system and dedicated AI detection for maximum accuracy

fix: add missing os import in backend utils

fix: bundle deberta tokenizer files to bypass Hub parsing errors

fix: apply DeBERTa tokenizer stability fix (use_fast=False) and protobuf

feat: initial clean project setup (root-level)