[ODE] Triangle-box collider contribution

[John Miles]

The biggest problem (really, the only problem) I've had with Erwin's current
rewritten solution is fall-through of (mostly) small, fast-moving objects.
This can be seen in test_moving_trimesh, when boxes get stuck inside the
wireframe bunny, and in my own overhead-view engine, when free-falling
objects fall through the terrain mesh.

Neither Jeff's nor Asko's latest versions perform any better than Erwin's in
that respect.  In fact, my empirical observation is that they're both quite
a bit worse.

Even if any outstanding bugs are fixed, I'm pretty sure this problem *can't*
be solved in the general case without a swept-volume collision detection
method that maintains awareness of the moving object(s)'s position between
frames.  Fast timesteps help, obviously, but there are always going to be
cases where that's not helpful or appropriate.

I don't see a compelling reason to abandon Erwin's current implementation
just yet.  It won't make sense to replace his code until someone writes a
frame-coherent collider that isn't vulnerable to missed collisions.

-- john


> Cool... if people want to try this out and compare it with the
> current CVS version (NOTE: trimesh-box has already been rewritten
> once since 0.039, and I think we really need a 0.0391 to give it
> a fair trial before we replace it again) then this stands a better
> chance of hitting CVS.
>