Although c9fa73a379 made the problem worse. The incorrect to snap to
edges when one of their vertices is behind the viewer is much older.
The problem occurred because, in perspective mode, when we project a
vertex with negative "zfac", its projected value is undefined as it
depends on the position of the other vertex.
Thus, both the distance test and the lambda in these cases were wrong.
The solution was to calculate the snap point directly in 3D, using
`closest_ray_to_segment_v3` and avoiding projecting the edge.
efd3c4b3c9