pub fn quantize_model(model: &Model, bits: u32) -> Model
Simple weight quantization: clamp weights to int8 range then dequantize. This simulates the effect of lower-precision storage.