Add support for serverless inference

[shchur]

Issue #, if available:

Description of changes:

• Added inference_mode parameter on CloudPredictor.deploy, TimeSeriesFoundationModel.deploy, and SagemakerBackend.deploy to provision serverless endpoints (inference_mode="serverless") in addition to the default realtime ones
• Added inference_config: Optional[Dict[str, Any]] for mode-specific overrides, forwarded to sagemaker.serverless.ServerlessInferenceConfig; defaults to memory_size_in_mb=4096, max_concurrency=5
• Raise ValueError if instance_type is set together with inference_mode="serverless"
• predict(wait=False) now returns a JobPredictionFuture (with output_path, status(), result()) instead of None; @overload added so type checkers narrow predict(...) to pd.DataFrame and predict(..., wait=False) to JobPredictionFuture
• Renamed backend-internal serve_config → fm_serve_config
• Renamed tutorial files: autogluon-cloud.md → cloud-predictor.md, foundation_model.md → foundation-model.md
• Added "Choosing an Inference Option" section + Serverless Endpoint subsection to the FM tutorial
• Added unit tests pinning the dispatch contract between inference_mode and sagemaker.Model.deploy(...) kwargs, plus a CI-only integration test deploying Chronos-2 to a serverless endpoint

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