Add grouped lag, rolling and expanding features

[lcrmorin]

Hello I was interested in having this in a personnal project. I used Gen AI to make this PR and reviewed it manually.

---

Closes #668. Adds groupby and sort_by to the existing forecasting transformers,
so stores/products/assets in the same dataframe cannot share lagged or aggregated
values. Multiple group columns form a combined key. Date-based results align on
both that key and timestamp. Existing ungrouped behaviour is unchanged.

This proposal follows the maintainer's invitation to supersede #727 with simpler
pandas grouped operations. It retains a unique row index; repeated dates across
groups can live in sort_by. No training history is carried into later transforms.
Per-variable window dictionaries are outside this change.

Validation: 143 forecasting tests pass on both pandas 3.0/sklearn 1.9 and pandas
2.2/sklearn 1.4, including independent-series reference comparisons, time windows,
frequency shifts, combined keys, drop_na and target alignment. User guides and
API docs updated; type and style checks pass.

Dependency requirements are unchanged. Testing covers the two configurations above, not the full supported-version matrix.