introduce and use maxContactCapacity (needs to be fixed in various other contact kernels)
implement sphere versus trimesh
disable new/sequential GPU batching (only uses 1 thread in a warp, slow but works on NVIDIA/Apple OpenCL)
y() -> y or getY() or [1]
z() -> z or getZ() or [2]
w() -> w or getW() or [3]
make sphere-convex and sphere-compound collision work (still issues remaining)
