Add optional `insample` parameter to `inverse_transform()` of data transformers and pipeline for simpler `Diff` handling

[dennisbader]

Inverse-transforming forecasts with the Diff transformer is tedious. It requires prepending the transformed insample target series to the forecasts for inverse_transform() to work.

We could add an optional parameter insample: TimeSeriesLike | None = None to inverse_transform() of InvertibleDataTransformer and Pipeline, that would simplify this process for the user.

This parameter would at the moment only be used by the Diff transformer and ignored by others.

Some requirements:

• The insample time series must be the transformed target series returned by Diff().fit_transform(series)
• Time frame:
  • It must start at the same time as the transformed target series
  • It must extend to at least one step before forecast start, but it can extend further into the future.
• Similar to how we handle insample for metrics such as mase(), we would extract the relevant parts of the series and prepend it the (forecast) series pass to Diff.inverse_transform()

Expected user flow:

from darts.dataprocessing.transformers import Diff
from darts.datasets import AirPassengersDataset
from darts.models import NaiveSeasonal

# `Diff()` or `Pipeline([Diff()])`
tf = Diff()
series = AirPassengersDataset().load()

series_tf = tf.fit_transform(series)

pred_tf = NaiveSeasonal(K=12).fit(series).predict(n=12)

pred = tf.inverse_transform(pred_tf, insample=series_tf)

That means that the data transformer handling in historical forecasts should also always pass the insample series.