Landmark bugs

[michaz]

Minimal fix for the two failing tests in #1687.

The first was already discussed (missing setter for the 'fallback' approximator).

The second one was due to an incorrect treatment of "minned out" "delta" weights. See comment in code.

[michaz]

I understand you are already investigating creating an extra factor for the delta weights, doing performance tests etc., but IMO step one is maintaining correctness, in this case doing the right thing when a delta weight is minned out. Then, step two is seeing how we can make sure this happens less often (if necessary).

I also removed the setMaxWeight from the test. PrepareLandmarks guesses this value by itself, and this must always work, i.e. it must guess in the right direction.

[michaz]

I think (?) one issue here was the test -- there even was a "setGet"-test, but it was rather technical and white-box -- it was basically testing that the bits would be twiddled in the same way the code says they would be twiddled. :-) But the important property -- that the remaining weight must never be over-estimated -- was not tested.

[karussell]

Thanks!