Add fixed point corrector and use it

* It is the same corr_ll layout, but with fixed point
* Also align matmul output to DAC units (provided that matrix coeffs are
 in DAC/nm, signed fixed point with 30 decimals)