[ODE] collision bug found!

[Erwin de Vries]

> i have just found a nasty bug in the collision code. erwin - this is
> most likely the source of the GeomGroup problems you were having.

You're right. That was the bug. I just checked it, and it works fine now.
Group-group collisions are now correct (this was probably caused by the
small bug i reported earlier), and there is no more corruption in the
geompointers. Excellent.

On a related note, i was thinking about the maximum of 255 contacts per
object. Is this enough? Shouldnt we raise this to 65536 for safety reasons?
I see problems arising when doing collision on groups containing other
groups whose might contain groups and a trianglelist containing complex
geometry. 255 contacts is a pretty small number in those cases.

On another related note. I figured out some interesting way for the
trianglecollider to make use of hierarchical temporal coherence when doing
collision on groups without the use of the aabb_test. Simply create a
trianglelist-geomgroup collision function which gets the aabb of the group,
and uses that to create its startingnodes in the tree. When the group
contains another group the temporal coherence function can use a stack-like
approach to push and pop temporal coherence info. There are probably
situations where this can lead to dramatic speedups.

Erwin