Fix perplexity computation, MQA/GQA models & models requiring position_ids

[fxmarty]

As per title

[fxmarty]

@Giuseppe5, with this, comparing to Brevitas model_eval I do get better eval results using CUDA_VISIBLE_DEVICES=3 python quantize_llm.py --fuse-sequences (for brevitas, simply using ppl = model_eval(model, validation_dataset, args.seqlen)):

Computing perplexity...: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 128/128 [00:02<00:00, 52.39it/s]
Perplexity (original model): 68.6707534790039
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 128/128 [00:01<00:00, 72.13it/s]
brevitas ppl: tensor(80.3409, device='cuda:0')

Which kind of makes sense as in optimum-amd a minimum context length is enforced.