Created by: jczaja

This PR is removing all tensor modifications. When doing this PR there were conflicts in Concat MKLDNN op , as this op started to use new API introduced by tensor modifications. Hence @xiaolil1 Could you please take a look at relevant changes.

As for performance results. We tested (so far, only resnet50) on CLX this PR vs this PR with older MKL-DNN. PR with older MKL-DNN shows that FPS was restored to level matching one before regressions. As for this PR (including MKL-DNN 0.18) performance is poorer by ~3 FPS (120 -> 117 ,resnet50 Fp32).

@xiaolil1 Please check this PR for performance and let us know your findings.

On making this revert. I used "git revert" on three commits that introduced tensor modifications. Manually I had to change:

• resolve conflict in concat mkldnn op
• Third commit was fixing and enabling transformer for mkl-dnn. This enabling of unit test was not reverted , while fix was reverted as it was related to tensor modifications.
• Transpose MKLDNN op was updated with labeling output tensor as kNCHW layout , to avoid having mkldnn_blocked_format.