Quantize weights independently along a selected channel axis, then dequantize with channel-specific scales.
