Suppprt static groups and fix bug

[qwopqwop200]

Adds static groups, a new feature of GPTQ.
Also fix disable exllama not working on fused attention.

[qwopqwop200]

I will be benchmarking llama soon. wait

[qwopqwop200]

All quantized from wikitext2.
Experiment with gptq code

LLaMA2-7B	Bits	group-size	Wikitext2	PTB	C4
GPTQ	4	128	5.7229	nan	7.1896
GPTQ(static groups)	4	128	5.7343	15849	7.1757
GPTQ(act-order)	4	128	5.5847	54.5089	7.1423
GPTQ(act-order&static groups)	4	128	5.5802	41.3790	7.1391

Experiment with AutoGPTQ code

LLaMA2-7B	Bits	group-size	Wikitext2
GPTQ	4	128	5.2595
GPTQ(static groups)	4	128	5.2692
GPTQ(act-order)	4	128	5.2388
GPTQ(act-order&static groups)	4	128	5.2542

[qwopqwop200]

PanQiWei
Experiments with static groups all seem to be ok. Now I think can merge them.

[PanQiWei]

Great work! Thank you so much for this fast implementation!