dotsocr 多并发下输出结果乱码

[test_dots.zip](https://github.com/user-attachments/files/30769679/test_dots.zip)

test_dots/
├── image.png # 原始输入图 (825x378)
├── call_dots_ocr.py # 参考脚本 —— 单服务并发压测
├── benchmark_compare.py # 对比脚本 —— A/B 两台服务器并发 QPS 对比
└── output_20260803_095654/ # 本次运行的全部产出

call_dots_ocr.py 单并发无问题
benchmark_compare.py 多并发会有乱码问题

当前测试的版本 0.21.0-b2

测试方法
python3 benchmark_compare.py \
--image image.png \
--url-a http://127.0.0.1:8000/v1/chat/completions \
--url-b http://127.0.0.1:8000/v1/chat/completions \
--model dotsocr \
--concurrency 1 4 8 16 32 \
--total 20

早期版本0.10.0-b4，客户反馈测试是正常的

Start with the attached call_dots_ocr.py and benchmark_compare.py scripts, using image.png and the provided endpoint command to reproduce the difference between single and multi-concurrency runs. Compare outputs at concurrency levels 1, 4, 8, 16, and 32, then trace the dotsocr serving path responsible for concurrent responses. Done means multi-concurrency output is no longer garbled and the comparison remains valid.

Written by the indexing model from the issue text.