Feature: Add sound libraries to dataset

[torridgristle]

The intention is for the model to learn things like engine roaring or sink draining and use the characteristics of these in a song.

Sound libraries like The General Series 6000 Sound Effects Library have very descriptive names for their sounds thst are like captions.

Perhaps by prefixing the non-song elements in the dataset with "sfx" (or something else if that's not a single token) it won't impact the creation of rhythm.