Linear optimization improvements

[tpitkanen]

Issue: #193

• improved optimization accuracy: numpy.polyfit is far better than scipy.interpolate.interp1d
• improved optimization stability: doesn't need monotonous sampled values anymore
• prominences are used for statistical weights instead of as cut-off values
• added user-selectable fitting polynomial degree

Done (from task list):

• _generate_mev_to_nm_function(): handle the case where there are really short peaks at surface or between valid peaks (ignore missing peaks or use the longest streak of non-zeroish peaks)
• selectable number of iterations

Future TODO:

• multi-thread initial espe runs (ElementSimulation.calculate_espe() uses multiple recoils for some reason, including self.optimization_recoils[0]. Check if a single recoil can be safely used instead.)
• remember which combination of NSGA-II/Linear and recoil/fluence is selected (not as simple as saving the other settings because the settings are saved using PropertySavingWidget, which doesn't work well on individual variables). As a quick fix, linear optimization could be the new default selection.
• add an option to get the starting solution from user
• unit tests
• add support for fluence optimization (currently only recoil optimization is supported)
• customizable shape (number of peaks, rectangular or not, peak at surface or deeper)