Fix streaming validation dataset causing infinite loop

[vominh1919]

Fixes #42

Problem

In train_diloco_torch.py, the validation dataset was loaded as part of a streaming (IterableDataset) load. The evaluate_model function iterates over eval_dataloader with a for loop, which never terminates because streaming datasets have no __len__ and never signal end-of-data.

Fix

• Removed validation from the streaming load_dataset call (training only)
• Load the validation dataset separately with streaming=False only when eval_steps is not None
• Apply the same tokenization and collation to the non-streaming validation set

This keeps training efficient with streaming while ensuring validation has a finite, known length.