Replace fp32 with int8-quantized model.onnx (390MB->98MB, accuracy within noise)

Upload README.md with huggingface_hub

v2: fine-tune from uitnlp/visobert (VN-native backbone) - beats v1 on every threshold

Upload README.md with huggingface_hub

Vietnamese toxicity model v2 (ONNX) - continue fine-tune from multilingual-toxic-xlm-roberta

initial commit