[CPU] Introduce SubModel op and Composite node

[EgorDuplensky]

with the idea of wrapping up (grouping) parts of the model / graph into a inner model / graph without any (almost) runtime overhead. SubModel op and Compoisite node are expected to perform no extra logic and to only execute an inner model / graph

Details:

• The naming "Composite node and SubModel operation" are discussable. A convenient option for the Composite node would be "SubGraph", but it is taken by snippets, which seems logical as well.
• the Composite node is not yet used by plugin
• single layer tests are to be added later. The problem is that currently it is impossible to use cpu plugin specific transformations and operations in scope of cpu func tests.

Tickets:

• ticket-id

[EgorDuplensky]

@maxnick Could you please take a look

[github-actions[bot]]

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

[EgorDuplensky]

@maxnick Could you please take a look one more time? Many changes have been done after an internal discussion, so there is no extra commit on top of the previous one, I have just overridden it.

[github-actions[bot]]

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

[maxnick]

Should we introduce functional / unit tests at this point?

[EgorDuplensky]

Should we introduce functional / unit tests at this point?

Please check a PR description with the reason why the tests are not added yet. Actually, this issue should be resolved if we decide to move a SubModel into internal opset, which is supposed to happen I guess, since GPU is also planning to re-use a tensor parallel logic.