Multi-GPU support for training and inference

[mshannon-sil]

In order to use larger models such as the 3.3B NLLB model, we would need to use two 40GB A100 GPUs. And for those using GPUs that have less memory than the A100, they would need multiple of them to have enough memory to run the 1.3B NLLB model. This means we'll need to add support for multi-GPU training and inference. Experiments should detect and use as many GPUs as are available.