see https://gitlab.com/sissopp_developers/sissopp
commits: up to 763e40e4deaf7a45dba3f0d8052cca053c1fa957

make it always consistent with the C++ calcs

Something is wrong with sharkML in some enviroments

1) Width reduced to 1e-5
2) SVM done with a C=100
3) SVM done with Kernel SVM but with a linear Kernel
4) Fixes to how algorithm breaks ties
5) Added a python SVM function
6) Modified Model python interface to accomadate SVM functions

Projection was only using last residual
Model did not print proper SVM misclassified

Debug tomorrow

performs better than Fortran in terms of accuracy
SVM is a challenge in terms of timing

use a linear prgramming method to check the feasabilty of a convex
combination of the classes to get an estimate of the number of points
in overlapping convex hull.

Compiles but not tested