Unify RoPE strategies #388

There has been some effort in hf/transformers to extract and share RoPE types across models, so we should be able to unify more, especially the loading/conversion parts. Some of the improvements are still in progress. A few relevant PRs:

• Llama: RoPE refactor huggingface/transformers#32135
• RoPE: model-agnostic RoPE refactor huggingface/transformers#31999
• Rename Phi-3 rope scaling type huggingface/transformers#31436