Minimum Smoothening Parameter bounds with H matrix

[vsl366]

The _P() Method already builds a lam * P matrix for each term. For the H matrix ( lower bounds on smoothening, we code a function build_minimum_penalties similar to build_penalties, which uses a new parameter lam_min (default 0) instead of lam. This term lam_min is added to every Term and init function, and is used to build the H matrix, using the _H() function. Previously, a snippet in line 750 already asks us to add the H matrix. Now, it is implemented as a function H, which returns the matrix, therefore it is changed to S += self._H(). This simple change enables us to add a minimum smoothening bound and now enforces it in the form:

GAM(s(0, lam_min=0.1) + s(1, lam_min=10)).fit(X, y)
enforces λ₁ ≥ 0.1, λ₂ ≥ 10

Closes #26