src: cpu: jit_bnorm_s8: change compute strategy

* Process single C cache-line over all spatial loop, instead of
* processing over all C loop, computing v = scale * v + shift every time

Author: dzarukin