* scripts to int8 quantize the thing
* target bf16 to uint8, 2x reduction
* able to load the model
* quantized working
* remove unused scripts
* conditional init depending on quantized
- Fixes RelativeUrlWithoutBase error when downloading models from fresh cache
- Resolves issue with kyutai/stt-1b-en_fr-candle model download
- Ensures proper URL handling in Hugging Face Hub interactions