Use pre-built FA2, vllm, quantization kernels in the dockerfiles

[fxmarty]

Feature request

We could just use

RUN pip install whatever_package.whl

which would tremendously speedup the build given that the target is always Hopper, Ampere.

This would involve having for example our own wheel index, handle the build when we want to upgrade, and use whl in TGI Dockerfiles.

Motivation

Building TGI is prohibitively slow for developers

Your contribution

I could work on that sometime