This fixes bug T68144.
The main issue with T68144 was that the ray to determine the number of intersections with the cylinder hit the cylinder exactly on an edge. I expect additional problems caused by floating point precision limitations. To solve both issues, I repeat the point-in-object test with at most 5 different rays as long as the result is uncertain. If all outcomes are uncertain, the majority vote is returned. All this should have little effect on performance, because uncertain outcomes should be unlikely.