[FEA] Lanczos Solver SM algorithm #2705
Description
Follow up to #2628. During the #2628 work the SM (which='SM' smallest magnitude eigenvalues) tests did not consistently pass, so it was omitted from that PR. For example one of the eigenvalues did not match in the output. SM is not required for the spectral embedding algorithm.
The cupy implementation also does not implement SM as they too were having issues with tests cases passing. See this thread cupy/cupy#4692
A user suggested using the shift-inverse method could help cupy/cupy#4692 (comment).
I have already implemented this in python in our test_sparse.py pytests to prototype and it does seem to help pass the tests and be more consistent. However, it does seem to require a non-trivial amount of effort since we need to implement a LinearOperator input type in cpp layer as well.