Added training with custom data support, fixed bugs with ddpm

[wu-ming233]

This pull request entails:

• Streamlining training on new data directly from midi files using an updated dataloader
• Supporting an arbitrary amount of tracks in the input, as opposed to the fixed (melody, bridge, piano) shape from POP909
• Extending the chord extractor with the option to round the duration of chords to the nearest beat. This significantly reduces error rate in preparing training data, while maintaining a reasonable accuracy
• Fixing collate_fn shifting unevenly in dataloader
• Fixing vanilla ddpm model not being able to train with PolyffusionLearner
• Fixing ddpm not being able to inference when length=1
• Enabling retraining from checkpoint with an increased max_epochs
• Implementing several convenient changes for batch inferencing

These changes are mainly just small additions to enhance user experience, and are completely backward-compatible.

[wu-ming233]

@aik2mlj The new commit implements the changes requested above, as well as two minor things:

1. When preparing data from MIDI, we would like to discard the non-melody parts, including drum tracks and melody tracks that have instrument program number >= 113. I've added this option in prepare_data.py, and it's enabled by default.
2. Unconditional vanilla ddpm lacked a way to specify which weights to use under the same checkpoint. I've added that, in a similar fashion as what's done in the other ldm models.

Let me know if anything else needs to be changed!

[aik2mlj]

Looks good to me! Thank you for your efforts. Now it's only the augmentation range [-6, 7) that needs to be fixed.