fix(tsfm): allowlist the transformers and estimators recipes use

[ShuxinLin]

Description

Fixes the allowlist issue from the #633 review (comment on resolver.py:33). The allowlist has no sktime transformers and no regressors, so these always fail with ... is not an allowed estimator class:

• the forecasting transforms block in run_recipe
• tabular sktime_class transforms in run_tabular_recipe
• the tsfm_regression tabular task

recipe_template advertises all three, and they worked before #633.

08e817b fixed this, but ebd792c reverted it, so the #633 head is back to the failing state. This PR covers only the allowlist. The other two review points (open-then-verify reads, confining training_args.output_dir) are not part of it.

Fix Details

substrate/resolver.py adds 33 exact paths to ALLOWED_CLASSES in three groups:

• run_recipe transforms, plus the classes they nest: Detrender, Deseasonalizer, ConditionalDeseasonalizer, Differencer, LogTransformer, BoxCoxTransformer, ExponentTransformer, SqrtTransformer, Imputer, HampelFilter, TabularToSeriesAdaptor, PolynomialTrendForecaster, TrendForecaster, StandardScaler, MinMaxScaler, RobustScaler
• run_tabular_recipe transforms: SummaryTransformer, Catch22, Rocket, MiniRocket, MiniRocketMultivariate
• run_tabular_recipe estimators: RandomForestRegressor, GradientBoostingRegressor, ExtraTreesRegressor, LinearRegression, Ridge, Lasso, KNeighborsRegressor, ExtraTreesClassifier, GradientBoostingClassifier, LogisticRegression, KNeighborsClassifier, KMeans

None of these takes a file path or a pickle. Some take a callable (KNeighbors* metric/weights, ConditionalDeseasonalizer seasonality_test, KMeans init). A JSON recipe can only fill those with plain values or with instances _build makes from allowlisted classes, and none of those instances is callable. A new test checks that for every allowlisted class. FunctionTransformer stays out.

Impact on Benchmarking

• No change to baselines: this restores estimators that worked before Aafeedback changes tsfm tools #633, and the code-execution routes stay closed.

Related Issues

• Refs: Aafeedback changes tsfm tools #633

Verification Steps

1. uv run pytest src/servers/tsfm/tests -k "not integration": 149 passed (145 before plus 4 new). All the existing code-execution tests still pass.
2. New tests in test_no_code_exec.py:
   • test_every_allowed_class_imports_and_is_not_callable: imports every path, which catches typos or sktime path drift, and checks that no instance is callable.
   • test_forecasting_transforms_still_run, test_tabular_regression_still_runs, test_tabular_sktime_transform_still_runs: one per broken route.
3. Manually, each of the 33 added classes succeeded end to end through run_recipe / run_tabular_recipe (32/32 runs; the scalers and trend forecasters run nested via _target_).
4. uv run pytest src/ -k "not integration": 769 passed. The 2 failures in src/observability/tests/test_file_exporter.py fail the same way on the Aafeedback changes tsfm tools #633 head without this change.

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules (no new findings on the changed files).
• I have signed off my commits (DCO).