Release inference code for OpenBEATs

[Shikhar-S]

Two new features in VERSA:

1. Sound class prediction from a fine-tuned OpenBEATs checkpoint.
2. Sound embedding extraction from pre-trained or fine-tuned checkpoint.

[ftshijt]

Thanks a lot for the great contribution. May I ask how you plan to store the embedding space in versa?

[Shikhar-S]

@ftshijt As discussed, added changes for

1. Storing embeddings as npy files
2. Similarity computation with a reference audio
3. Class prediction to output class names and log probabilty.

[ftshijt]

Thanks a lot. I think the interface is good to go. Before further steps, I would like to check the following items with you first:

• I'm okay with directly putting the architecture in versa, but it might be easier if we simply utilize it from espnet since we do have the dependency there? Mostly, I want to make it connect to your future versions (if any), which might reduce your effort to push the updated code twice. (It's up to you~)
• We definitely want to keep the options of using local checkpoints, but at the same time, do you mind smoothing the usage with automatic download of the model (e.g., from huggingface etc.)?
• For the class prediction cases, it would be super helpful if you could provide more examples with different downstream tasks~