lm-eval polishing and speed-up

[12010486]

This pull request updates the Habana model adapter in examples/text-generation/model_adapter.py to improve device handling, logging, and padding logic for text generation on HPU devices.

More importantly, it speeds up both loglikelihood and generate_until based tasks at 0% change in accuracy.

Some examples:

PT_HPU_LAZY_MODE=1  python run_lm_eval.py --model_name_or_path meta-llama/Llama-3.2-1B-instruct \
--attn_softmax_bf16 --use_hpu_graphs --limit_hpu_graphs   --use_kv_cache --bf16 --sdp_on_bf16 --trim_logits \
--batch_size=32 --tasks gsm8k_cot_llama -o eval_gsm8k.json --num_fewshot=8 \
--fewshot_as_multiturn --apply_chat_template True 

Improves throughput from 8.57it/s to 20.83it/s

PT_HPU_LAZY_MODE=1  python run_lm_eval.py --model_name_or_path meta-llama/Llama-3.2-1B-instruct --attn_softmax_bf16 --use_hpu_graphs --limit_hpu_graphs   --use_kv_cache --bf16 --sdp_on_bf16 --trim_logits --batch_size=16 --tasks ifeval -o eval_ifeval.json --num_fewshot=0 --fewshot_as_multiturn --apply_chat_template True

Improves throughput from 2.73it/s to 3.39it/s

PT_HPU_LAZY_MODE=1  python run_lm_eval.py --model_name_or_path meta-llama/Llama-3.2-1B-instruct --attn_softmax_bf16 --use_hpu_graphs --limit_hpu_graphs   --use_kv_cache --bf16 --sdp_on_bf16 --trim_logits --batch_size=32 --tasks hellaswag -o eval_hellaswag.json

Improves throughput from 46.36it/s to 198.44it/s

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[regisss]

LGTM