Feature (quantizer): Adding weight normalization-based integer quantization

[i-colbert]

Adding experimental narrow per-channel weight normalization-based signed integer quantizer based on Quantized Neural Networks for Low-Precision Accumulation with Guaranteed Overflow Avoidance, by I. Colbert, A. Pappalardo, and J. Petri-Koenig, with support for both L1 and L2 weight normalization.

Using the decoupled rescaling integer quantization arithmetic where the weight normalization statistics calculation (d_w) and norm vector parameterization (g) are combined with the scaling factor to become the pre-clipping scaling factor (i.e., pre_scale) and the conventional scaling factor (s) is the post-clipping scaling factor (i.e., post_scale). For further details on the arithmetic, see ParameterPreScalingWeightNorm. For further details on the weight normalization-based quantization technique, see the referenced paper.