WIP: Implement LTU model

[The-Mats]

We want to experiment with LTU as it is a promising model that combines natural language and audio classification to create a more capable and diverse model. Gong et al. created a newer version of the model named LTU-AS, which uses whisper features and performs better especially for speech and music. I think we should first try LTU and then maybe LTU-AS!

For inference they provide shell scripts that locally load a website for easy interaction, which we don't really need.

The following steps are needed to integrate the model into the BS pipeline:

• Read paper
• Download the needed model files and there are several as we also need the LLM (Not sure if we can remove it later)
• Add missing packages
• Create a model file that includes the needed code from their inference_gradio.py
• Add get_embedding()
• Create experiment and model config
• Test with BEANS
  • linear probing
  • few shot
  • fine tuning ?
• Test with BirdSet
  • linear probing
  • few shot
  • finetuning ? is that possible?
• ...