Drop GPTQ support

[carmocca]

Closes #582
Closes #583

GPTQ is inference only, requires a conversion step, and the implementation we use is much slower than bitsandbytes. The only upside is that it uses less memory at inference time.

There's a lot of research happening around inference quantization and having this implementation in the repo is not worth it anymore.

[Andrei-Aksionov]

Sorry for procrastinating with my AutoGPTQ integration, but I've started working on it (again) and there should be something to look at somewhere next week.
So the question is whether you are dropping GPTQ (and its variants) or just the current implementation?
If just the current implementation, then this PR doesn't close #583.

[carmocca]

The current implementation, however, as long as any new additions are not as useful and usable as existing implementations (for now it's only bnb) then we wouldn't be interested in adding them.

This PR should close #583. I suggest opening new issues proposing the addition of new techniques.

For instance https://github.com/IST-DASLab/marlin was released yesterday and includes its own GPTQ implementation. Perhaps AutoGPTQ is no longer the best alternative.

edit: they'll be adding marlin support with AutoGPTQ/AutoGPTQ#514

[Andrei-Aksionov]

Great!
Then I'll continue with AutoGPTQ.