Check if there are degenerate eigenvalues in svd

[eperim]

If one tries to compute the sensitivity of svd with a matrix that has two or more identical eigenvalues (like the identity, for instance), values will blow up at this line https://github.com/invenia/Nabla.jl/blob/master/src/sensitivities/linalg/factorization/svd.jl#L68

It would be nice to have a check and then either warn or fail gracefully in case that happens.

PS: This seems to be an issue with pytorch and autograd as well.

[oxinabox]

Does this tie to:
W Wang et al ‎2019: Backpropagation-Friendly Eigendecomposition
https://papers.nips.cc/paper/8579-backpropagation-friendly-eigendecomposition.pdf

[eperim]

That's awesome. The issue seems to be exactly that.