Created by: jczaja

Our Validation detected that training performance is significantly decreased after introducing: #21466 . Problem will be properly fix When Grad NHWC is supported. But since FWD of NHWC is ported to 1.6.2 but not Grad NHWC then it is advisible to restore training performance without waiting for proper fix. Changes are restoring implementation of Grad of conv MKLDNN kernel like it was before #21466.