Unfortunately, Pythagorus and binary don't play well together, so rounding
errors are inevetible when testing with a slope. However, 1e-6 seems to be
a good epsilon (printf's %g hides it nicely :).
This one demonstrates the need for more information in the bsp tree
(surface polygons). When the box collides with a corner where one side is
flat and the other angled, but there's a partition plane cutting the two,
the box can instead collide with the angled side before it hits the corner.
However, there is still a problem with moves where the box is always cut
by the plane: both sides need to be tested (done), but when the first side
checked allows longer motion than the second, but still collides, only the
first side is checked. The shorter motion is required.
I'm not sure the end point needs to be moved at all, but I'll leave it
alone for now. I have a couple of failing test cases that seem to be caused
by not handling moves where the box is always cut by the plane.
Rather than setting allsolid when the trace fails to leave solid space,
clear it when the trace enters non-solid space. This is necessary because
the trace might visit only one node and thus the failure to leave solid
space will not be detected. This fixes the problem with hipnotic's bobbing
water.
