ScalarQuantizer

Scalar quantization and codebook lookup for TurboQuant.

After random rotation spreads quantization error uniformly, scalar quantization maps each element independently to an N-bit integer code. This is simpler and faster than vector quantization while achieving good quality thanks to the rotation preprocessing.

The quantizer uses a uniform symmetric scheme:

• Compute per-group scale = max(abs(group)) / ((2^(bits-1)) - 1)

• Quantize: code = round(value / scale), clamped to -2^(bits-1)+1, 2^(bits-1)-1

• Dequantize: value ≈ code * scale

Groups of 32 elements share a single FP16 scale factor.