[timeseries] Accept seasonal_period in TimeSeriesScorer.__call__ with a deprecation warning - #5957

Issue #, if available: none

Description of changes:

#5084 (horizon weights) moved both prediction_length and seasonal_period from TimeSeriesScorer.__call__ to the constructor. Passing prediction_length to __call__ still works and emits a FutureWarning. Passing seasonal_period there, as code written for earlier versions does, now fails with an error that doesn't point at the cause:

from autogluon.timeseries.metrics import MASE

MASE(prediction_length=5)(data, predictions, seasonal_period=3)
# TypeError: MASE.save_past_metrics() got multiple values for keyword argument 'seasonal_period'

This happens because seasonal_period stays in **kwargs and reaches save_past_metrics() next to the explicit seasonal_period= argument.

This PR handles it the same way as prediction_length: it emits a FutureWarning ("deprecated and will be removed in v2.0"), sets self.seasonal_period, and removes the key from kwargs. The result is the same as constructing the metric with seasonal_period=....

Tests:

• New test_when_seasonal_period_passed_to_call_then_warning_is_raised_and_value_matches_init for MASE, RMSSE and SQL. It checks the warning and that the value equals the one from seasonal_period in the constructor. DUMMY_TS_DATAFRAME is hourly (default seasonality 24), so the test also fails if the argument were dropped instead of applied. On master it fails with the TypeError above.
• timeseries/tests/unittests/test_metrics.py: 268 passed. ruff check and ruff format --check pass.

While checking this I also compared MAE, MSE, RMSE, MAPE, SMAPE, WAPE, MASE, RMSSE, WQL, SQL and MQL against their documented formulas on series with missing past values and a seasonal period of 3, and all of them matched.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.