[docs] LLM inference

[stevhliu]

This PR creates a lightweight LLM inference optimization guide focused on just providing a brief explanation and copy/paste-able code examples so users can instantly get going without necessarily needing to get too in-depth in the underlying conceptual details about how it works. It includes:

• static kv-cache + torch.compile
• speculative and prompt lookup decoding
• attention optimizations (FlashAttention-2 and SDPA)
• quantization

[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.

[stevhliu]

For each of the topics covered, I was wondering if:

• @ArthurZucker, would you mind having a look at the static kv + torch.compile section?
• @gante, would you mind having a look at the speculative decoding section?
• @younesbelkada or @SunMarc, would you mind having a look at the attention optimization and quantization sections?

[SunMarc]

Thanks for the great work @stevhliu ! I left a couple of suggestions.

[ArthurZucker]

Thanks!

[gante]

Very cool 🙌

[younesbelkada]

Fantastic work ! 🚀

[amyeroberts]

Looks great - thanks for adding!

[ArthurZucker]

Sorry for the late review here!