app: improve end point detection for smart colorization.
Previous algorithm was relying on strokes of small radius to detect points of interest. In order to work with various sizes of strokes, we were computing an approximate median stroke thickness, then using this median value to erode the binary line art. Unfortunately this was not working that well for very fat strokes, and also it was potentially opening holes in the line art. These holes were usually filled back later during the spline and segment creations. Yet it could not be totally assured, and we had some experience where color filling would leak out of line art zones without any holes from the start (which is the opposite of where this new feature is supposed to go)! This updated code computes instead some radius estimate for every border point of strokes, and the detection of end points uses this information of local thickness. Using local approximation is obviously much more accurate than a single thickness approximation for the whole drawing, while not making the processing slower (in particular since we got rid of the quite expensive erosion step). This fixes the aforementionned issues (i.e. work better with fat strokes and do not create invisible holes in closed lines), and also is not subject to the problem of mistakenly increasing median radius when you color fill in sample merge mode (i.e. using also the color data in the input)! Also it is algorithmically less intensive, which is obviously very good. This new version of the algorithm is a reimplementation in GIMP of new code by Sébastien Fourey and David Tschumperlé, as a result of our many discussions and tests with the previous algorithm. Note that we had various tests, experiments and propositions to try and improve these issues. Skeletonization was evoked, but would have been most likely much slower. Simpler erosion based solely on local radius was also a possibility but it may have created too much noise (skeleton barbs), with high curvature, hence may have created too many new artificial endpoints. This new version also creates more endpoints though (and does not seem to lose any previously detected endpoints), which may be a bit annoying yet acceptable with the new bucket fill stroking interaction. In any case, on simple examples, it seems to do the job quite well.