Regularization towards initial guess

[nicolearetz]

Made changes to method_of_snapshots:

• For some non-Euclidean inner products, the computed reduced spaces weren't orthonormal (the multiplication of the states with the eigenvalues and eigenvectors is too sensitive to errors). Fixed it through additional Gram-Schmidt steps.
• Found an error where the computed reduced space would have shape (n, r) with r > n. This would lead to issues with the orthonormality check. Fixed it so that at most n eigenvalues are treated as non-zero (which is how many there should be at most anyway)

Made changes to PolynomialOperator:

• Added parametric version
• changed the computation of the condensed Kronecker product so that it doesn't compute the whole Kronecker product first. This speeds-up the ROM computation time a lot.

Made changes to models (mono):

• Added the possibility to specify an initial guess to the fit function. The regularized OpInf problem is then regularized towards the initial guess instead of zero.
• added tests

[shanemcq18]

I think initial_guess needs to be part of the solver, not the model, or the ROM class won't see it.

[shanemcq18]

Should we get rid of these print() calls? Or do we want to keep them?

[shanemcq18]

Should the initial_guess be part of the solver instead of the model? So, something like this:

rom = opinf.ROM(
    model = opinf.models.ContinuousModel(
        operators="AB",
        solver=opinf.lstsq.L2Solver(reg, initial_guess=something),
    )
)

I think it should be this because the initial guess is only used for Tikhonov-type solvers, right? Doing it this way would also play nicely with the ROM class, otherwise you can't specify the initial_guess unless you are calling the model's fit() method directly.