ENH overlap-add convolutions

I think this is standard practice for 1D signals, because they often have to deal with very long signals wrt filter size, but this can be applied across the board: Imagine a very big 2D image, or, more realistically, a 3D array which is actually a video of 2D frames which you want to analyze with spatiotemporal gabors. This will require overlap-add in the time axis.

So the general setup might be something where you specify the axes that you want to have overlap-add in, and it could be a WaveletTransform or OverlapAddConvolution module that takes care of it.

[janden]

The other potential advantage here, from what I understand, is that it may give us a way to compute filters independent of image size, since we’re storing them in the spatial/time domain.

[eickenberg]

Working with 1D data of shape (384, 300000) is showing me the limitations of not having this implemented. My GPU maxes out in memory at (30, 300000), which, if we made this into one long line, would correspond to like 400s of mono audio at 22050. It might be nice to be able to go a bit above that without having to manually chunk data etc