Training examples use the supported Python package and strategy

The Python RL example installs nodus-compute. The fine-tuning reference uses DDP for its multi-node example; managed distributed training rejects DeepSpeed ZeRO strategies. Distributed training remains Beta and requires qualified capacity.