Add support for dimension selection for layer normalization

[WilliamTambellini]

Summary

Add support for dimension selection for layer normalization
Atm the normalization axis is imposed to be the last logical dim:
https://oneapi-src.github.io/oneDNN/dev_guide_layer_normalization.html#difference-between-forward-training-and-forward-inference

Problem statement

Cannot select which axis to normalize on.

Preferred solution

a new runtime param for the norm prim execution.

We are open to prepare a PR if ok.

[vpirogov]

@WilliamTambellini, thanks for the offer. Considering that this feature involves API changes an RFC describing the problem statement and API changes would be great starting point. An implementation is welcome as well.