[BUG] Potential sequence length bug in preprocessed record

[Chetansahney]

Describe the bug

In pytorch_forecasting/data/_tslib_data_module.py, _preprocess_data() sets:

"length": len(series)

But series is a dict-like record, so this returns the number of keys, not the sequence length.
This can produce incorrect metadata for downstream consumers that rely on length.

To Reproduce

import pandas as pd
import numpy as np

from pytorch_forecasting.data import TimeSeries
from pytorch_forecasting.data._tslib_data_module import TslibDataModule

#simple toy series
df = pd.DataFrame(
    {
        "group": np.repeat([0], 20),
        "time": pd.date_range("2020-01-01", periods=20),
        "target": np.random.randn(20),
        "feature": np.random.randn(20),
    }
)

ts = TimeSeries(
    data=df,
    time="time",
    target="target",
    group=["group"],
    num=["feature"],
)

dm = TslibDataModule(
    time_series_dataset=ts,
    context_length=8,
    prediction_length=4,
    batch_size=2,
)
dm.setup()

idx = dm._train_indices[0]
processed = dm._preprocess_data(idx)
original = ts[idx]

print("processed length:", processed["length"])
print("expected length:", len(original["t"]))

Expected behavior

processed["length"] should equal the actual time-series length (e.g., len(original["t"])), not the number of keys in the series record.

Additional context

This can cause subtle bugs where sequence-dependent logic reads a wrong length value from preprocessed records.

Versions

pytorch-forecasting: 1.7.0
Python: 3.12.x
PyTorch: 2.x
OS: Linux

[phoeenniixx]

FYI @PranavBhatP