[E2E][Accuracy] `google/gemma-2-2b` and `meta-llama/Llama-3.2-1B` Huggingface models are broken

[anmyachev]

CI ref: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/18809058806/job/53681691151#step:7:379

[anmyachev]

Fixed in #5399

[kwasd]

Thank you!