(Hopefully) corrected bug in when searching for possible translation.

When the (new) RndWlk routine is not successful in suggesting translation,
the system uses the (old) routine, which samples around the cell and then
chooses among the free positions. However, the sampling was is "binned"
and so not every pixel within the "bin" in the sampled mask is guaranteed
to represent for sure a non-colliding translation. The bug was that the
system had not enough trials at hand when choosing from this sampled mask,
and if it was unlucky enough, it kept choosing always wrong pixels...

M    src/cellScm.cpp
M    src/toolbox/cell_movement.cpp