lot of changes that can be summarize in "able to stich more than two images" (modulo a bug in the initialization of the optimization)

In more details :
- matching algorithms based on "groups match", faster and better results
- improvement in the harris detection, remove the threshold (or more like use a smaller value... like 0.0) on cornerness, this increase the number of interest point, and split the layer in zones to increase the distribution on the image
- introduction of the notion of models which are in charge of feeding the optimizer, and then output the transformation that can be feed to the blender - plus countless of changes to remove the limitation to two images