Fix GeometryFixer via buffer-0 noding robustness check

[dr-jts]

This PR fixes the GeometryFixer failure cases identified in #852.

The GeometryFixer problem is that invalid hole rings in the input are fixed via buffer(0), but the output of buffer(0) is still invalid. This causes the union of the "fixed" holes to fail with a TopologyException.

The fundamental problem is that the fast noding used in the buffer algorithm is non-robust, and thus can produce incorrect topology. In the Overlay code this is handled by checking the correctness of the noding after it is created. This is effective, but does impact performance. This is not done in the buffer algorithm, probably because buffer topology building catches noding errors later on. It hasn't caused any known problems - until this issue.

The fix is to check the fast noding done in buffer. If this check fails, the buffer code retries using a more robust noding strategy. To avoid impacting performance for other buffers the noding check is done only for distance=0.

This fix also contains an improvement to CascadedPolygonUnion to switch to using OverlayNG as the robustness fallback, since the test case was affected by the buffer changes. (This code path will soon be obsolete anyway, once OverlayNG becomes the default.)

NOTE: for some reason GEOS works for these test cases, so this fix does not need to be ported.

Signed-off-by: Martin Davis [email protected]