Created by: bingyanghuang

This PR is to refactor the PR #16342, one PR #18570 for matmul kernel has been merged. This PR contains:

• Pass for modifying the graph

• A fix for the merged matmul kernel for supporting last dimension value is not equal

This PR can be run with the ERNIE, and there is a little accuracy diff with this PASS are under fixing.

Pass for graph modification:

• Before

• After