Created by: jczaja

This PR introduces changes to performance issue described in #16248 (closed) . Changes here Are adapting mkl-dnn ops to tensor modifications so overhead by setting legacy function: set_format() is removed. Also set_format() is removed .

@xiaolil1 If you could run same performance experiement (just Int8 part) as You did in #16248 (closed) It would be helpful confirmation that modifications are going into good direction.