Better Call SAL: Towards Learning to Segment Anything in Lidar