Module quantizer 

Groupwise symmetric quantization with f16 scales.

For each group of group_len values:

• scale = max(|v_i|) / qmax
• q_i = round(v_i / scale), clamped to [-qmax, +qmax]
• u_i = q_i + qmax (bias to unsigned for packing)

Functions

compute_scales

Compute f16 group scales for a frame.

dequantize

Dequantize packed codes using f16 scales, writing f32 values.

dequantize_f32

Dequantize packed codes using f32 scales, writing f32 values.

frame_fits_scales

Check if a frame fits within existing f16 scales (within drift tolerance).

frame_fits_scales_f32

Check if a frame fits within existing scales (within drift tolerance).

quantize_and_pack

Quantize a frame using pre-computed f16 scales and pack into bitstream.

quantize_and_pack_f32

Quantize a frame using pre-computed f32 scales and pack into bitstream.

scales_to_f32

Pre-convert f16 scales to f32 for hot-path use.