[BUG] save_time_based_splits function does not support CPU mode well

Bug description

Function save_time_based_splits in data_utils.py does not support CPU mode correctly. In particular, function _save_time_based_splits_cpu assumes using Rapids libraries, moreover Dask Dataframe seems incorrectly imported.

Steps/Code to reproduce bug

Using code from examples, just with option CPU set to True (https://github.com/NVIDIA-Merlin/Transformers4Rec/blob/main/examples/getting-started-session-based/01-ETL-with-NVTabular.ipynb)

sessions_gdf = df.read_parquet(BASE_PATH / "processed_nvt/part_0.parquet")
from transformers4rec.utils.data_utils import save_time_based_splits

save_time_based_splits(
data=nvt.Dataset(sessions_gdf),
output_dir=BASE_PATH / f"session_by_day",
partition_col="day-first",
timestamp_col="session_id",
cpu=True
)

Expected behavior

No exception is thrown and data are splitted.

Environment details

• Transformers4Rec version: 23.12.0
• Platform: Ubuntu 20.04.3 LTS
• Python version: 3.8.10
• Huggingface Transformers version: 4.30.2
• PyTorch version (GPU?): 2.4.1
• Tensorflow version (GPU?): 2.7.0

Additional context

Start in data_utils.py at save_time_based_splits and _save_time_based_splits_cpu, then reproduce the issue using the CPU=True example from the linked notebook. Done means the function raises no exception and splits the data as expected.