rfcs: extend layer norm to support root mean square normalization

[mzhukova]

Proposal to extend current Layer Normalization primitive to support RMS normalization via new flag.

Read here: https://github.com/uxlfoundation/oneDNN/blob/rfcs/rfcs/20251004-rms-norm/README.md
Implementation: #3068
JIRA: https://jira.devtools.intel.com/browse/MFDNN-13287

// Note: Current proposal is aligned with keras RMS normalization, but not Layer Normalization with rms_scaling due to keras-team/keras#21234.

[mzhukova]

also inviting @pengzhao-intel and @gaurides to take a look

[mzhukova]

confirmed with @pengzhao-intel and @gaurides, that this options is preferred --

(Preferred) Omit mean in the output:
Pros: Cleaner design and avoids unnecessary memory requirement.
Cons: Requires users to adjust their code to handle the absence of the mean for RMSNorm comparing to LayerNorm.

[mgouicem]

Regarding open questions:

• I would vote for dnnl_use_zero_mean or dnnl_no_mean, as this is closer to how we split dnnl_use_scale and dnnl_use_shift.
• I would vote for omitting mean, but it might be worth checking with PyTorch team what the expectation is.

[mzhukova]

hi @dzarukin, @mgouicem and @gaurides, thank you for your feedback!
I believe I've addressed all the current comments, so could you please re-review/close opened threads and let me know if anything else is needed here or approve.

[gaurides]

LGTM