add gpt_neox support

[twaka]

Hi, I'm trying to support gpt_neox architecture.

• Support for gpt-neox model #41

awq/models/gpt_neox.py results in almost the same code with bloom and I didn't try to scale fused query_key_value.
Also I conducted eval.py with gpt_neox variants because of not enough resource to run original gpt-neox-20b.
Much appreciaed if someone can evaluate with it.

rinna/bilingual-gpt-neox-4b

before

Task	Version	Metric	Value		Stderr
wikitext	1	word_perplexity	15.0848
		byte_perplexity	1.6611
		bits_per_byte	0.7321

after

Task	Version	Metric	Value		Stderr
wikitext	1	word_perplexity	15.8949
		byte_perplexity	1.6774
		bits_per_byte	0.7462

stabilityai/stablelm-base-alpha-7b

before

Task	Version	Metric	Value		Stderr
wikitext	1	word_perplexity	17.1947
		byte_perplexity	1.7023
		bits_per_byte	0.7674

after

Task	Version	Metric	Value		Stderr
wikitext	1	word_perplexity	17.6954
		byte_perplexity	1.7114
		bits_per_byte	0.7752

[casper-hansen]

Looks good to me! Thanks for implementing this architecture.