MODEL: Add SqueakOut

[NickleDave]

https://www.biorxiv.org/content/10.1101/2024.04.19.590368v1.full.pdf
https://github.com/gumadeiras/squeakout

A couple things that will require extra work for this model:

• It's trained as an autoencoder, on images. So prep will need to make images, and then the loss will be a regression loss (I think MSE). They evaluate with Dice. So we don't get a direct comparison with e.g. binary classifiers very easily.
• It's not explained in the paper how images are prepared. I believe it's this code in VocalMat:
  https://github.com/ahof1704/VocalMat/blob/5be886dc611e90e41a9b8c2ee428a69c9d6ebcf0/vocalmat_classifier/vocalmat_classifier.m#L257

        img = imresize(flipud(mat2gray(A_total(:,T_min:T_max))),size_spectrogram);
        img = cat(3, img, img, img);
        imwrite(img,fullfile(vpathname, vfilename, 'All', [num2str(id_vocal)  '.png']))

The idea of saving spectrograms as png files of grayscale images is interesting. I wonder if it takes up less space than an npz file because of the image compression.

When using models that take standard images as input, we do increase parameters a bit by requiring the model to have 3 channels in the input layer.