speculative decoding complete guide added

[shirinyamani]

What does this PR do?

In this PR, I've provided a comprehensive guide on Speculative Decoding techniques aimed at enhancing inference speed. It includes a basic implementation of the core ideas from the paper Accelerating Large Language Model Decoding with Speculative Sampling, applied to GPT-2. Alongside the implementation, I've added detailed explanations of the underlying mathematical principles and an intuitive breakdown of why speculative decoding works.

This PR serves as a complementary extension to the speculation.md file in the TGI repository, offering further practical and theoretical insights into speculative sampling.

[Narsil]

Hi !

Thanks a lot for this contribution. This seems like a great contribution, but not really on point with text-generation-inference core goals. We do not provide speculation with draft models (for various reasons) for instance, nor is code really wanted in documentation (at least not this amount of code).

It feels to me like this could become a blog of your own, that we could maybe link to explain what is speculation ? As for TGI's documentation, we don't think it's our documentation's job to explain speculation (there are vastly better ressources out there) but more to explain HOW to leverage speculation within TGI, maybe a few unique tradeoffs we made, and only reference external content for core comprehension of the technique.

Does that make sense to you ? Would linking to a blog of your own work for you ?

[Narsil]

Closing for now, feel free to reopen.