Support sound-source separation at inference time

[sammlapp]

Using a google-research product https://github.com/google-research/chirp/tree/main/chirp/birb_sep_paper#separation

the approach looks like

• load both the separation model and the classification model
• perform source separation during inference, separating audio into channels
• predict on each channel

[sammlapp]

seems like it would requires supporting TensorFlow model inference ( #817 )

[sammlapp]

We now support MIxIT separation model (TensorFlow Hub model) in the Bioacoustics Model Zoo, but its unclear how to efficiently incorporate that model into an inference workflow with Pytorch-based models