Add fit_predict / fit_predict_proba to TabularCloudPredictor

[shchur]

Fuse fit + batch predict into a single SageMaker training job: fit on train_data and predict on a separate test_data inside the training container (predict_after_fit=True), mirroring the existing TimeSeries fit_predict path. Avoids a second cold start, data upload, and the predictor-tarball round-trip of a separate batch-transform job.

• train.py: branch the predict_after_fit block on predictor_type; tabular reads a new test_data channel and writes the full [pred, proba...] frame (regression: just [pred]). Raises NotImplementedError on image columns.
• TabularSagemakerBackend: fit override that validates test_data covers the train feature columns client-side before launch, then injects it as a data channel.
• TabularCloudPredictor: fit_predict -> pd.Series and fit_predict_proba(include_predict) -> (pred, proba) | proba, matching the predict / predict_proba split. wait=False returns None; fetch later via get_fit_predict_results / get_fit_predict_proba_results.
• Tests: pure-unit wiring/validation tests plus a parametrized classification/regression integration test.

Issue #, if available:

Description of changes:

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