[CPU] Fixed first token latency for compressed models

[dmitry-gorokhov]

Details:

• This PRs restore oneDNN IP primitive loop order for weights compression case to fix 1st inference latency degradation
• OneDNN PR: [FORK][FIX] Restore choose_loop_order logic for weights decompression case oneDNN#254

Tickets:

• CVS-145808

[github-actions]

This PR will be closed in a week because of 2 weeks of no activity.

[wenjiew]

@yuxu42 Let us monitor this PR with the latest oneDNN 3.8 integration to evaluate whether this is needed. If so, let us see how to complete the PR. Thanks!

[moslex]

Original ticket is closed. 145808
Can we close this PR ?

[yuxu42]

closing the PR, as the perf improvement for weight compressed models have been implemented by other PRs, such as #28576