Avoid saving predictors for FoundationModel prediction jobs

[shchur]

Summary

• add a save_predictor training flag that defaults to True
• set save_predictor=False only for tabular and time-series FoundationModel prediction jobs
• preserve existing fit_predict predictor-artifact behavior
• stage in-job predictions in /tmp to avoid duplicating them in SageMaker output.tar.gz
• disable unused leaderboard generation for tabular foundation-model prediction
• verify any reported SageMaker model artifact is either absent or contains no predictor files

Testing

• Validated through the PR integration checks; final rerun pending.

[github-actions]

Job PR-275-0fe116d is done.
Docs are uploaded to https://d12sc05jpx1wj5.cloudfront.net/PR-275/0fe116d/index.html