Add time series converter for regression forecasting

[Irozuku]

Summary

Adds TimeSeriesWindowConverter, which reshapes a forecasting dataset into the supervised rows a regression model needs. Second of four branches implementing forecasting.

A forecasting dataset is a date column and a value column: one row per point in time, with nothing to regress on. The converter turns it into lag_k, ..., lag_1, target, where each row carries the k values that came before target. That is a plain tabular regression problem, so the result is used with the existing RegressionTask and the regressors DashAI already has, rather than needing anything new.

Given a window of 3 and the series 100, 120, 115, 140, 150, 160:

lag_3,lag_2,lag_1,target
100,120,115,140
120,115,140,150
115,140,150,160

---

Type of Change

Check all that apply like this [x]:

• Backend change
• Frontend change
• CI / Workflow change
• Build / Packaging change
• Bug fix
• Documentation

---

Changes (by file)

• DashAI/back/converters/simple_converters/time_series_window.py: new converter. SUPERVISED = True, since the series to window is the target rather than something in scope, so it arrives as y. CHANGES_ROW_COUNT = True, since the return value replaces the whole dataset, which is what lets the date column disappear and the lag columns be the entire output. Scope accepts Date only; window_size is the single parameter.
• DashAI/back/initial_components.py: register the converter.
• tests/back/converters/test_time_series_window.py: new. 12 tests, the first being the example above asserted cell for cell.

---

Testing (optional)

• Apply the converter to a time series dataset and train a regression task.

---

Notes (optional)

Two consequences are documented in the component description, since neither is obvious from the UI:

• Everything outside the scope and the target is discarded, because the windowed rows no longer correspond to the original ones.
• Consecutive rows overlap by k - 1 values, so a shuffling splitter puts near duplicate rows on both sides of a split and reports a score that is too good. Split chronologically.

Rows are sorted by date before windowing, so source ordering does not matter. Refused: repeated dates, a missing or non-numeric target, a scope that is not exactly one Date column, and a window leaving no complete row. Irregular spacing warns rather than blocks, since a lag then covers a different span for different rows, but monthly data with 28, 30 and 31 day steps is legitimate.

Multi horizon targets are out of scope: RegressionTask has outputs_cardinality = 1, so they would have nowhere to go until a multi output regression task exists.