swr: allocate all scratch space in one go for vertex buffers