Created by: wozna

This PR adds a squash dequantize-scale pattern. That squash removes scale operator and rewrites that scale to dequantize operator scale. Only works when the bias of the scale op is 0. There are 12 such pairs in the Ernie model.