feat: enable Hugging Face eval_loss on sharded datasets

[DanielAsadi]

Summary

Hugging Face eval_strategy is documented on GR00T, but the sharded dataset path cannot actually produce eval_loss:

• DatasetFactory.build() asserts eval_strategy == "no" and always returns eval_dataset=None.
• The mixture iterator is infinite, so an eval pass would never finish.
• HF Accelerate wraps the eval iterable dataloader and broadcasts CPU objects through NCCL.
• Trainer.prediction_step drops the loss because Gr00tN1d7 has empty label_names and no return_loss argument, so eval emits eval_runtime but no eval_loss.

This change keeps training behavior when eval_strategy is "no". When it is not:

• split episodes deterministically into disjoint train/eval sets
• iterate eval shards once, padding the schedule to world_size * num_workers
• return a raw eval DataLoader (same as train) and put the processor in eval mode
• always return the model forward loss from prediction_step

Test plan

• tests/gr00t/data/test_dataset_factory.py
• tests/gr00t/data/test_sharded_datasets.py
• tests/gr00t/experiment/test_trainer_eval_dataloader.py
• tests/gr00t/experiment/test_trainer_prediction_step_loss.py
• 1-GPU fine-tune smoke: eval ran at steps 10 and 20, eval_loss appeared in trainer state / logs, training resumed, checkpoint written
• 4-GPU smoke still pending (exercises shard padding / multi-rank collectives)