Sharded fine-tune path cannot emit Hugging Face eval_loss #778

eval_strategy is accepted on TrainingConfig, but the sharded dataset path cannot actually produce eval_loss.

1. DatasetFactory.build() asserts eval_strategy == "no" and always returns eval_dataset=None.
2. ShardedMixtureDataset iterates forever, so HF eval would never finish.
3. Accelerate prepares the eval iterable dataloader and broadcasts CPU objects through NCCL.
4. Even if eval runs, Trainer.prediction_step drops the loss: Gr00tN1d7 has empty label_names and no return_loss argument, so logs show eval_runtime only.

Training with eval_strategy: "no" is unaffected.