feat: Add support for phi4

[jlonge4]

This PR adds support for Meta's Phi-4 model by adapting the existing LLaMA implementation.

The Phi-4 architecture follows the LLaMA architecture closely, with the main difference being in how the weights are stored (fused qkv_proj and gate_up vs separate projections).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[dacorvo]

@jlonge4 thank you very much for this pull-request: adding support for phi4 would be awesome.

We are however heavily refactoring the export mechanism to remove the dependency to transformers-neuronx and simplify the contribution of new models.

Can you take a look at that pull-request and see if it would make it easier for you to add support for phi4 based on the new HLO backend ?

[jlonge4]

Hi there @dacorvo , just took a look at the difference and it certainly seems a lot slimmer! I think my effort would be the same in regard to the most important part for this which is the load_weights function. However it would obviously get rid of a lot of boiler plate. I am down to rework this PR and merge into add_hlo branch if you prefer.