Rather than separate chips for training and inference, Tesla is converging on a single family of chips — powering cars and robots individually while scaling into massive clusters for training in data centers.