Rework stats, scale, threshold and zero-point calculations

[nickfraser]

Different modules seem to have different implicit assumptions around stats, scale, threshold - e.g., whether they are represented in the log2 domain, whether the are expressed as thresholds or scale factors.

It might be worth giving this a look with fresh eyes and seeing if we can apply the same explicit assumptions everywhere. It might not be possible without incurring some other undesired complexity elsewhere.