Explain data parallelism, pipeline parallelism, and tensor parallelism, and when you would use each at Mistral AI to scale training.
