Sphere fit eigensolver

[ThibaultLejemble]

I believe there is an error in SphereFitImpl::finalize() when solving the eigenvalue problem:
https://github.com/poncateam/ponca/blob/d2cda8378fcc27c8f27be1f94bca85f90e35796a/Ponca/src/Fitting/sphereFit.hpp#L68C28-L76C28.

The vector u = [uc ul^T uq]^T of the sphere is the eigenvector associated with the minimal eigenvalue of the general eigenvalue problem Au = lambda Cu (in the code A = m_matA and C = matC).

The method computes M = inv(C) A to solve the standard eigenvalue problem Mu = lambda u by using the self-adjoint eigensolver on M^T M. However, I think that the eigenvectors of M and M^T M are not the same (they are if M is symmetric, which is not the case here).

I propose to use the Eigen::GeneralizedEigenSolver to solve Au = Bu and then take the real part of u (I think the imaginary part is null since A and C are symmetric).
I quickly tested it, and it gives visual results that are slightly better (when I look at points projected onto their locally fitted sphere).

[nmellado]

Thanks a lot.
Seems to be fine.
To be able to review it, could you please add mathematical details in the user documentation associated to this function ?