Delay SVD of the transformation matrix in TauSampling and MatsubaraSampling

[shinaoka]

The SVD can be done when the fit is called first time.

[SamuelBadr]

I implemented this in the lazy-svd branch but had to make the *Sampling types mutable, so I'm a little hesitant to merge into main. If it degrades performance, maybe the current solution (specifying factorize = false via kwarg in the *Sampling constructors) is preferable. What do you think?

[shinaoka]

maybe the current solution (specifying factorize = false via kwarg in the *Sampling constructors) is preferable.

Sorry, I was not aware of this change!
OK, let us stick to your solution.

BTW, we should add some description on this point in the "sparse-ir-tutorial" since this is crucial when evaluating an object on a large number of frequencies and tau points.