Made a few updates to align AMLB with AutoGluon 1.0's API.

Note that all changes are backwards compatible. For AutoGluon 1.0 exclusive logic, I added if/else checks so it continues to work with older versions: Further, the mainline AMLB code works with AutoGluon 1.0 already as old methods were deprecated but not removed, it will just log deprecation warnings.

Changes related to AutoGluon 1.0:

1. persist_models -> persist
2. persist max_memory=0.4 -> removed as it is now the default
3. predictor.predict_from_proba to allow for decision threshold calibration (not enabled by default)
4. Remove silent=True to leaderboard call, as it is now silent by default.

General improvements unrelated to AutoGluon 1.0:

1. Allowed config.max_runtime_seconds to be None when good_quality or high_quality preset is specified and avoid crashing.
2. set -e: Ensure install script crashes when an error occurs rather than continuing (previously could lead to corrupted installs sneaking by and being ran on datasets, now they should error properly)
3. Switch repo location from awslabs/autogluon to autogluon/autogluon (it was migrated last year, awslabs/autogluon redirects to autogluon/autogluon)