refactor CPU llama inference code

[faaany]

What does this PR do?

This PR refactors the current CPU llama inference code to make code clean. The major changes are as follows:

• introduce a new class _IPEXLlamaAttention and move the attention-related OPs and attention forward code to _IPEXLlamaAttention
• introduce a new class _IPEXLlamaMLP and move the MLP-related OPs and forward code to _IPEXLlamaMLP
• simplify _patch_llama_model
• rename _IPEXLlamaDecoderLayerRef to _IPEXLlamaDecoderLayer
• refactor the forward mtehod of _IPEXLlamaAttention into gemm, rope and sdpa

Please note that this PR is based on the unmerged PR #725 by Jiqing as can be seen in the commit history.

[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.

[echarlaix]

https://github.com/huggingface/optimum-intel/pull/725 is now merged, you mind rebasing @faaany ?

[faaany]

725 is now merged, you mind rebasing @faaany ?

Cool! Rebase done, pls have a review. Thx!

[faaany]

Hi @echarlaix , I manually-checked the changes in #725 and fixed the bugs introduced through rebasing. Now all tests passed, I think we are good to go.