Bug: Training script with rm-pad-id will only train on subset of dataset.

[TueVNguyen]

If we use the rmpad_with_pos_ids or rmpad mode, we will train on only a subset of the full dataset.
This is due to the behavior of the UnpackDataCollator() function, which selects only the first sample in each training batch.

Please see the comment in the script: test_rm_pad_dataset.py

---

# Before fix the bug, the result is:
# Total samples in the data loader: 1226.0
# Total samples in the dataset: 1000
# Total unique ids got from dataset: 248

# After fix the bug, the result is:
# Total samples in the data loader: 1237.0
# Total samples in the dataset: 1000
# Total unique ids  got from dataset: 992