Consider switching to Separating Axis Theorem for collision checks

[davesmith00000]

Narrow phase bounds checks are currently done using different methods for the two shapes, and special cased when the two collide (which won't scale).

Since all our shapes are axis aligned, generalising to use SAT collision checks should be pretty straightforward, and I think it will give us the restitution values for free, regarding the other open ticket around box corner collisions.