Initialization strategies for multi-component models

[luisfabib]

This PR introduces new strategies to initialize the non-linear parameters of the N-component models in the fitmultimodel function.

Three strategies are introduced:

• Spreading: For each N-component model, the non-linear parameters start values are spread equidistantly over the box constraints. The number of components are changed in a forward matter, i.e. 1,2,...,N. Previously the default in fitmultimodel.
• Splitting: For each N-component model, the non-linear parameters start values are selected by splitting the location and spread of the components obtained from the N-1 component fit. The number of components are changed in a forward matter, i.e. 1,2,...,N. Now the default in fitmultimodel, faster and more accurate than the spreading strategy.
• Merging: For each N-component model, the non-linear parameters start values are selected by merging the location and spread of the components obtained from the N+1 component fit. The number of components are changed in a backward matter, i.e. N,N-1,...,1. More accurate than the spreading strategy but might be slower than the splitting strategy.

[stestoll]

What is the reason you apply these initialization strategies to the spread? I would think it is needed only for the location. Otherwise, what would you do about other basis function parameters such as asymmetry parameters?

Second question: do all dd model functions have a single location parameter, or there any that have two? This would have implications here.