Add flag to prevent duplicate inference and refactor PyTorch hub config

[edwinzhng]

• For the inference client, I refactored it to avoid duplicate detection using a flag.
• For hubconf.py I refactored it so that all of the model loading logic is offloaded to helper functions

Now the steps to add a new pretrained model are:

1. Upload the model weights + settings to the howl-models repo and create a new release
2. Add a short function to load the model in hubconf.py similar to the one for hey_fire_fox

   def hey_fire_fox(pretrained=True, **kwargs):
   """Pretrained model for Firefox Voice"""
   engine, ctx = _load_model(pretrained, "res8", "howl/hey-fire-fox", **kwargs)
   return engine, ctx

[edwinzhng]

@ljj7975 should I add the instructions for uploading a new pretrained model into the README?