Support wavs+csv labels for building small classifiers

[sdenton4]

We currently use a folder-of-folder format for labeled data. If we allow providing a directory of audio files and a CSV of labels, then we will be able to support:

a) Evaluation with published fully-annotated datasets,
b) Simpler import of people's pre-existing labeled data,
c) Multi-label audio examples,
d) 'Strong negative' labels (via an extra CSV column)