- InStatePath::solve
    BiRRT* does not work properly at all:
        - discontinuities due to an algorithmic mistake involving qProj_
        - not using path projectors, while it should
    DiffusingPlanner does not work properly sometimes
        - Some collisions were detected on paths solves for romeo-placard
          both with discrete and continuous (Progressive, 0.2) path validations


- InStatePath::mergeLeafRoadmapWith
    this is inefficient because the roadmap recomputes the connected
    component at every step. A better merge function should be added to roadmap.cc


- path optimizations after solving once :
    They are done "locally" (for each of the leafs separately) in C++ with a
    hard-coded optimizer (Graph-RandomShortcut) and then globally with the
    command ps.addPathOptimizer("Graph-RandomShortcut"). Due to how this works,
    this is like doing two times the same thing so the first should be removed.
    However bugs have been observed when the global optimization is used (Core
    dumped). For unknown reasons, in GraphOptimizer::optimize, the dynamic
    cast of current->constraints() may fail on *some* 0-length paths. Ignoring
    empty paths (the "if" I have added) seems to make the problem disappear
    but the reason why some 0-length paths don't have the right Constraint type
    is still to be found.