Fix for eos in eager mode

[12010486]

A fast repro is as below:

PT_HPU_LAZY_MODE=0 python run_generation.py --model_name_or_path meta-llama/Llama-3.2-1B-Instruct --use_kv_cache --max_new_tokens 100 --do_sample --batch_size 2 --prompt "Hello world" "How are you?" --sdp_on_bf16 --no-ignore_eos

Which would have failed without this PR.

We get:

Input/outputs:
input 1: ('Hello world',)
output 1.1: ("Hello world! This is my first time using this platform. I am excited to share my knowledge and experiences with the world. I am a student of computer science and engineering, and I am currently working on my master's degree.\n\nMy area of interest is in artificial intelligence, machine learning, and deep learning. I have been studying these topics for several years and have gained a good understanding of the concepts and techniques involved. I have also been experimenting with various deep learning frameworks such as TensorFlow and PyTorch.\n\n",)

input 2: ('How are you?',)
output 2.1: ("How are you? Today is a beautiful day!\nI'm doing great, thanks for asking! It's a lovely day to be outside, isn't it? I love days like this, when the sun is shining and the birds are singing. There's something so uplifting about being in nature, don't you think?\n\nI've been meaning to get out and enjoy the great outdoors, but I've been stuck inside a lot lately. How about you? What are you up to today? Do you have any fun plans",)


Stats:
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Input tokens
Throughput (including tokenization) = 108.17886215980566 tokens/second
Average first token latency         = 22.46260239626281 ms
Average rest token latency          = 18.413477945447234 ms
Average end to end latency          = 1848.2638252025936 ms
Memory allocated                    = 5.62 GB
Max memory allocated                = 5.76 GB
Total memory available              = 94.62 GB
Graph compilation duration          = 5.586424403998535 seconds
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

--no-ignore_eos is just slightly slower than --ignore_eos in eager, 113.073 vs 108.179 tokens/second. This is expected

[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.

[yafshar]

LGTM!

[regisss]

LGTM