使用hifigan_24k报错 KeyError: 'model_state'

**Summary[问题简述（一句话）]**
Vocoder使用hifigan_24k报错 KeyError: 'model_state'，输出为杂音。使用pretrained或g_hifigan则正常。

**Env & To Reproduce[复现与环境]**
模型从readme的百度网盘下载
版本为最新的main分支。
从百度网盘下载模型放入对应目录后，vocoder选择hifigan_24k的时候就会报错。但是点击生成仍然可以正常运行，只是出来的都是杂音。

**Screenshots[截图（如有）]**
```
Traceback (most recent call last):
File "/Users/xxx/Files/MockingBird/toolbox/__init__.py", line 467, in init_vocoder
vocoder.load_model(model_fpath, model_config_fpath)
File "/Users/xxx/Files/MockingBird/vocoder/wavernn/inference.py", line 37, in load_model
_model.load_state_dict(checkpoint['model_state'])
KeyError: 'model_state'
```


Start in toolbox/__init__.py at init_vocoder and then read vocoder/wavernn/inference.py, especially load_model where checkpoint['model_state'] is accessed. Compare the hifigan_24k model downloaded from the README with the working pretrained or g_hifigan models. Done means hifigan_24k loads without the KeyError and generated audio is no longer noise.

Written by the indexing model from the issue text.