qwen2.5 output is not correct under dml ep

**Describe the bug**

model: Qwen2.5-7B-Instruct convert to onnx model using onnxruntime_genai.models.builder

when use cpu ep : the output is correct:

but the output of directml ep is abnormal:

**To Reproduce**
1. commoad covert model:
python -m onnxruntime_genai.models.builder -m "D:\models\Qwen2.5-7B-Instruct" -e dml -p int4 -o "D:\models\Qwen2.5-7B-Instruct-onnx-dml"
python -m onnxruntime_genai.models.builder -m "D:\models\Qwen2.5-7B-Instruct" -e cpu -p int4 -o "D:\models\Qwen2.5-7B-Instruct-onnx-cpu"

2. test python code:

```python

import onnxruntime_genai as og
from transformers import AutoTokenizer

model = og.Model("D:\models\Qwen2.5-7B-Instruct-onnx")
tokenizer = og.Tokenizer(model)
tokenizer_stream = tokenizer.create_stream()
# tokenizer = AutoTokenizer.from_pretrained("D:\models\Qwen2.5-7B-Instruct-onnx-dml")

# Set the max length to something sensible by default,
# since otherwise it will be set to the entire context length
search_options = {"max_length": 2048, 'do_sample': True, 'top_k': 40, 'temperature': 0.2, 'top_p': 0.8,
'repetition_penalty': 1.2}

chat_template = (
"\n<|im_start|>user\n{}<|im_end|>"
"\n<|im_start|>assistant\n"
)

prompt = f'{chat_template.format("tell a story about your favorite food!")}'

print(prompt)
input_tokens = tokenizer.encode(prompt)

params = og.GeneratorParams(model)
params.set_search_options(**search_options)
params.input_ids = input_tokens
generator = og.Generator(model, params)

print("Output: ", end='', flush=True)

import time

s = time.time()
n_tks = 0

try:
while not generator.is_done():
generator.compute_logits()
generator.generate_next_token()

new_token = generator.get_next_tokens()[0]
n_tks += 1
print(tokenizer_stream.decode(new_token), end='', flush=True)
e = time.time()
print(f"\n{n_tks / (e - s):.2f} tks/s")
except KeyboardInterrupt:
e = time.time()
print(f"\n{n_tks / (e - s):.2f} tks/s")
print(" --control+c pressed, aborting generation--")

print()
del generator

```

**env**
- intel Ultra 7 155H igpu , driver version : 32.0.101.5972
- OS: windows11 , python3.11 , onnxruntime-dml 1.19.2 , onnxruntime-genai-directml 0.5.0.rc1 ,

Start by reproducing the two `onnxruntime_genai.models.builder` conversions and the supplied Python generation script, comparing CPU and DirectML results. Investigate the DirectML path and the reported environment versions; done means Qwen2.5 generation produces correct output under the DirectML execution provider.

Written by the indexing model from the issue text.