[PyTorch] Enable generic QK norm support (+ RMSNorm/LayerNorm)

[negvet]

Description

For the training stabilization purposes, QK tensors might be normalized.

• Enables generic QK norm support.
• Enables RMSNorm/LayerNorm as a normalization types (in addition to L2Normalization).
• Enables qk normalization before and after RoPE (following both Qwen and Llama approaches).

Extention of #1864
Fixes #1958

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[Marks101]

Great, thanks for looking into this so quickly. This all looks good, but:

I have to admit that there is one more thing that I just noticed. The Qwen3 models apply QK normalization before RoPE, see here, this is in contrast to this implementation which is based on LLama4.
I was not aware that there are two different formulations for this. Sorry for that.

[negvet]

I have to admit that there is one more thing that I just noticed. The Qwen3 models apply QK normalization before RoPE, see here, this is in contrast to this implementation which is based on LLama4. I was not aware that there are two different formulations for this. Sorry for that.

This flexibility might be worth it; I supported it.
It brings some complexity and additional argument, but it is manageable, I guess.
Let's give it a review.

[negvet]

/te-ci pytorch

[Marks101]

Great, that was super quick! This looks very good to me 🥳

[negvet]

/te-ci pytorch

[timmoon10]

LGTM