在 bistream 2.0 模型中，零样本推理结果会将参考音频的结尾拼接到推理结果上，有时会出现跳过开头内容的问题。

Describe the bug As the title suggests, the zero_shot inference result using the 2.0 model under the main branch of bistream will slightly splice the end of the reference audio to the inference result (such as the manger, In the morning...); occasionally, there will also be a problem where the beginning is skipped, such as In the morning, the actual beginning is the morning.