add simple demo ppl test with wikitext2

[qwopqwop200]

This is a demo code using ppl to test the performance of AutoGPTQ.
opt - 125m
RTN baseline 37.27 ppl
GPTQ for llama 29.21 ppl
AutoGPTQ 29.87 ppl
Some performance degradation is observed, but it seems to be a minor error.

[PanQiWei]

Yeah, I can also confirm that there is a small gap between AutoGPTQ and GPTQ-for-LLaMa by running this example.

I'm not quite sure if it's because AutoGPTQ's logic used in model.quant() function and how attention_mask been processed is slightly different with GPTQ-for-LLaMa.

This finding is important, maybe there's room for improvement here.

I think this pr can be merged.