ENH: Add TCN models + DAS datasets

[NickleDave]

• Add DAS network architecture, converted to pytorch. From https://github.com/janclemenslab/das/blob/master/src/das/models.py
  • Will require implementing the DAS-specific vendored version of TCN:
  • As well as the vendored version of the Kapre layer:
  • Ideally should have torchvision-style builder functions for each model that can specifies the correct network parameters and loads corresponding weights, i.e. das_bengalese_finches
• For training, add DAS to the FrameClassificationModel family
• Add datasets, from https://data.goettingen-research-online.de/dataverse/das?q=&types=dataverses%3Adatasets&sort=dateSort&order=desc&page=2
  • Because these datasets are audio + annotations this will be one place to refactor VocalDataset as described in Refactor frame classification models to use single WindowedFramesDatapipe #574 ENH: add BaseVocalDataset that uses vocles #539 and use it as a base for these specific datasets
    • Will need to make sure item from these datasets matches what's returned by AudioSequence in DAS: https://github.com/janclemenslab/das/blob/9ea349f13bdde7f44ba0506c5601078b2617cb45/src/das/data.py#L81

[NickleDave]

Re-opening this because we haven't added datasets yet

[NickleDave]

Also because after discussion with @yardencsGitHub it's clear that I have made a mistake in my implementation and need to do further work on this

Leaving the initial implementation in this branch:
https://github.com/vocalpy/vak/tree/implement-DAS-model

Will remove from the main branch for now

Next step will be to compute segment error rate as we do using the original model in keras to understand what I am doing wrong

[NickleDave]

I think we can still add a TCN model, e.g. from here
https://github.com/vocalpy/Nicholson-Cohen-SfN-2023-poster
We just shouldn't claim to perfectly replicate DAS