[BUG] Discrepancies between Pipeline and AutoRegressivePipeline in handling a given ts after transforms

[d-a-bunin]

🐛 Bug Report

There is a difference in a way how Pipeline and AutoRegressivePipeline are handling a given ts in forecast in relation to transforms.

• AutoRegressivePipeline seems to take into account transforms which were already applied before calling forecast
• Pipeline seems not to take into account transforms which were already applied before calling forecast

Expected behavior

I think that the behavior should be the same. It isn't really obvious which is better, but probably Pipeline's.

How To Reproduce

from copy import deepcopy

from etna.datasets import TSDataset
from etna.datasets import generate_ar_df
from etna.pipeline import AutoRegressivePipeline, Pipeline
from etna.transforms import LagTransform, AddConstTransform
from etna.models import LinearMultiSegmentModel


def main():
    df = generate_ar_df(n_segments=3, start_time="2020-01-01", periods=100, freq="D")
    ts = TSDataset(df=df, freq="D")
    
    model = LinearMultiSegmentModel()
    transforms = [
        LagTransform(in_column="target", lags=[7, 8, 9, 10])
    ]
    autoreg_pipeline = AutoRegressivePipeline(model=model, transforms=transforms, horizon=7)
    pipeline = Pipeline(model=model, transforms=transforms, horizon=7)

    additional_const_transform = AddConstTransform(in_column="target", value=10)
    ts.transform(transforms=[additional_const_transform])

    autoreg_pipeline.fit(deepcopy(ts))
    pipeline.fit(deepcopy(ts))

    autoreg_forecast = autoreg_pipeline.forecast()
    forecast = pipeline.forecast()
    
    mean_autoreg = autoreg_forecast.to_pandas(features=["target"]).mean().mean()
    mean = forecast.to_pandas(features=["target"]).mean().mean()

    assert abs(mean_autoreg - mean) < 5


if __name__ == "__main__":
    main()

Environment

No response

Additional context

No response

Checklist

• Bug appears at the latest library version