Bugfixes and improvements for v0.5.0

[shchur]

Issue #, if available:

Description of changes:

1. Bugfix: save_path typo — parse_backend_predict_kwargs was reading kwargs.get("persist") instead of kwargs.get("save_path").
2. cloud_output_path resolution centralized — Extracted resolve_cloud_output_path() in aws_utils.py, shared by both CloudPredictor and FoundationModel. Supports fallback from ~/.autogluon/cloud.yaml config, warns when a prefix already has objects. Returns None when nothing is configured (defers error to backend property access), preserving backwards compatibility for flows that don't need it (attach_endpoint, etc.).
3. get_fit_predict_results reworked — Fetches ag_args.pkl from the SageMaker training job's input channel instead of local disk, so it works when the tmpdir is gone or the job is reattached in a fresh process. Uses download_file (single key) instead of download_data (prefix scan).
4. predictions_path validation moved to backend — Consolidated in SagemakerBackend.fit() so it's enforced regardless of entry point (TimeSeriesCloudPredictor or FoundationModel).
5. predictions_path param added to TimeSeriesFoundationModel.predict() — Lets users specify a custom output location.
6. SageMaker SDK version gate — import autogluon.cloud raises ImportError if sagemaker>=3.0 is installed.
7. Bucket validation in register() — Rejects bucket names containing prefixes (e.g. my-bucket/path), normalizes s3:// scheme and trailing slashes.

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

[github-actions]

Job PR-238-bbb6df4 is done.
Docs are uploaded to https://d12sc05jpx1wj5.cloudfront.net/PR-238/bbb6df4/index.html

[github-actions]

Job PR-238-d4afe6e is done.
Docs are uploaded to https://d12sc05jpx1wj5.cloudfront.net/PR-238/d4afe6e/index.html