Add new texture model

[billbrod]

This is a good issue for a summer intern or new grad student.

The PortillaSimoncelli texture model implementation is a bit finicky, because we're trying to match the matlab implementation, which is a bit finicky. Our general recommendation is that people should use a new implementation that captures the same statistics, but does so in a modern way. It should compute the cross- and auto-correlations across scale and orientation. One thing that might be tricky is the phase-doubling trick the authors use to make correlations work across scales.

[nikparth]

I'd say a big issue with the current implementation is the fact that in computing the various statistics there are a lot of places where for loops are done over all scales/orientations to compute statistics such as autocorrelation per band. To make this really scalable with batches, you would also need to figure out how to vectorize those operations. That would be primary focus since that would have the biggest impact on speed. In addition, adding the ability to process a batch will require careful stepping through the code and either modifying the operations to handle the batch dimesion or maybe utilizing the in-built torch vmap: https://pytorch.org/tutorials/prototype/vmap_recipe.html. Using vmap may be the cleanest way because we can still maintain the intuition of computing these statistics per image but then vmap that operation across the batch dimension.