[sycl-free-inference-for-llms] run llama3-8B with Pytorch for xpu and get the base line

[Dewei-Wang-sh]

measure each kernel's perf in eager mode and torch.compile

[vlad-penkin]

This ticket is part of umbrella ticket:

• [sycl-free-inference-for-llms] Port and evaluate LLama3-8B and Granite-8B #2170

[zhczhong]

• Enable the flaggem on triton-xpu backend and benchmark the performance on several oob model #2976
• Enable the liger-kernel on triton-xpu and benchmark the performance #2977