Increase flexibility for output of waveform-to-labels models

[cwitkowitz]

Waveform-to-label models have little flexibility when it comes to the output predictions. As an example, I am trying to export an automatic music transcription model which produces multiple labels for each time step. This is not possible with the current version of audacitorch, and there does not seem to be a straightforward workaround.