Multidimensional SVD

There notes seem relevant for you

But there is not yet support for batched svd in pytorch. It seems that numpy just do a for loop over the batches, which would be easy to add on the CPU side, but on the GPU side it would be trickier.