Unary patch embeddings

[brunosan]

We have a ticket when I open the issue of rescitting the neck of the Unet from a bundle of embeddings, one per band, to just one, and currently we can outpout the average, average per group, or each embedding. Keeping the raw embedding seemed too big, but this blog post on unary embeddings, suggest that transforming the raw patch embeddigns to unary migh be a new solution.

Filing to test.

They see a reduction of accuracy but also a vast increase in performance.