[Bullet] Rigid and Kinematic bodies use different collision detection algorithms.

[e344fde6bf]

Godot version:
3.2.3

OS/device including version:
N/A

Issue description:

From the bullet manual, the default algorithms used for collision detection are taken from this matrix:

	box	sphere	convex, cylinder, cone, capsule	compound	triangle_mesh
box	boxbox	spherebox	gjk	compound	concaveconvex
sphere	spherebox	spheresphere	gjk	compound	concaveconvex
convex, cylinder, cone, capsule	gjk	gjk	gjk or SAT	compound	concaveconvex
compound	compound	compound	compound	compound	compound
traingle_mesh	concaveconvex	concaveconvex	concaveconvex	compound	gimpact

In Godot's bullet backend rigid bodies seem to use the default algorithms, however kinematic bodies seem to use gjk for all convex shapes including box and sphere.

This can make a major a difference for boxbox collisions when using large collision margins. Bullet uses the margin value to round out sharp edges with spheres that have a radius equal to the margin value. However, bullet's boxbox algorithm is a special case that does not use the margin value, while gjk does.

Here's a demonstration showing the difference for the boxbox case. I used OpenSCAD's minkowski() to generate visual representations of the exact collision shapes used by bullet in the gjk case. The red boxes are rigid bodies, the gray ones kinematic:

bullet-boxbox-collisions.mp4

There doesn't seem to be much difference for the spheresphere and spherebox cases, however minor differences in how rigid and kinematic bodies calculate the separation between shapes can contribute to #31981 (though this is not the main cause of that bug).

Steps to reproduce:

• Create kinematic and rigid boxes with large margin values and collide them together.

Minimal reproduction project:
bullet-boxbox-collisions.zip

[Calinou]

cc @pouleyKetchoupp