Convert MusicLDM

[sanchit-gandhi]

What does this PR do?

Adds the conversion script for MusicLDM and a new pipeline class, closely based on the existing AudioLDM pipeline.

Changes compared to the existing AudioLDM pipeline:

1. AudioLDM only uses the CLAP text branch. MusicLDM uses the full CLAP model (text + audio branch) for similarity scoring: the cosine similarity is computed between the generated waveforms and the text inputs, and the audios ranked based on these scores (most similar -> least similar). For MusicLDM, this scoring has quite a big effect on the quality of the generated audios when num_waveforms_per_prompt>1.
2. Addition of the CLAP feature extractor for pre-processing the audio waveforms for the CLAP audio branch: the feature extractor is registered as a new module in the __init__, and is used in the score_waveforms method

TODO:

• Finalise design - are we happy with adding a new pipeline as described above, or do we want to try and make the existing pipeline compatible with the two changes described above, possibly at the expense of greater code complexity (need to condition every call to the text_encoder)
• Add tests & update docs

cc @Vaibhavs10 @sayakpaul

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sayakpaul]

Depending on the complexity, I think it's okay to add a separate pipeline for this.

[patrickvonplaten]

Very clean! Only left some nits