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)