fix(temporal): emit feature names in the order transform produces

[ChrisW09]

Fixes #15.

Problem

LagFeatureTransformer, RollingStatsTransformer and CyclicalTimeTransformer all build
their output by hstacking one (n_rows, n_features) block per lag / statistic / component,
so the columns are block-major. But they inherit
BasePreTabTransformer.get_feature_names_out, which walks
zip(input_features, self._output_sizes()) and emits feature-major names.

For any input with more than one column, the labels are simply wrong:

Lag  names: ['A_lag0', 'A_lag1', 'B_lag0', 'B_lag1']
Lag  row 0: [  1. 100.   0.   0.]     # col 1 = 100 is B's lag-1, labelled A_lag1
Roll row 0: [  1. 100.   2. 200.]     # actual order mean_A, mean_B, max_A, max_B

Single-column input happens to be correct, which is why the existing tests pass.

Fix

Override get_feature_names_out in each of the three classes to iterate blocks outermost,
matching the actual column layout.

Scope note

I kept the existing lag{j} / roll{j} / cyclic{j} suffix scheme rather than renaming to
_mean / _sin / _cos. The issue suggested that rename as a bonus and it is genuinely
more informative — but it changes public output names for single-feature users whose labels
are correct today, which is a different kind of change from fixing a mislabelling. Keeping
the suffixes means:

• single-feature output is byte-identical (['t_lag0', 't_lag1'] before and after);
• no existing test needed changing. My first pass did rename, and it broke three tests
  that encode the current scheme — a good signal it belongs in its own PR.

Happy to follow up with the rename if you want it.

Result

Lag  ['A_lag0', 'B_lag0', 'A_lag1', 'B_lag1'] | row0 [  1. 100.   0.   0.]
Roll ['A_roll0', 'B_roll0', 'A_roll1', 'B_roll1'] | row0 [  1. 100.   2. 200.]
Cyc  ['A_cyclic0', 'B_cyclic0', 'A_cyclic1', 'B_cyclic1']

Tests

Five added to tests/test_temporal.py, using a two-column fixture where B = 100 * A so
every column is identifiable from its value alone:

• one per transformer asserting names against the real column contents
• a parametrised check that default (None) names still work for single-feature input
• a width check that len(get_feature_names_out()) == transform(X).shape[1]

Full suite: 487 passed, 9 xfailed, no existing test modified. ruff check clean on changed
files; pyright unchanged at its pre-existing 71 errors.

🤖 Generated with Claude Code