Refactor generation benchmark to compare with AWQ and HQQ

[dacorvo]

The same generation metrics can now be evaluated on models quantized using quanto, BitdsAndBytes, hqq and AutoAWQ.

Example on princeton-nlp/Sheared-LLaMA-1.3B with int4 weights and float16 activations on an Nvidia A10:

library	perplexity	prediction	latency
fp16	8.85	0.83 %	24 ms
quanto	9.99	0.81 %	69 ms
bnb	9.35	0.82 %	35 ms
hqq	9.05	0.83 %	96 ms
AutoAWQ	9.02	0.19 %	7 ms