Optimize refit_folds logic

[Innixma]

Issue #, if available:

Description of changes:

• Optimize refit_folds logic
• This is an updated version of this PR: Tabular: Switched KNN and RF to use refit_folds instead of use_child_oof #1351
• Currently does not enable refit_folds=True for any model, but a follow-up PR will test enabling for KNN.
• refit_folds=True will fit a bagged ensemble of a given model and then immediately call refit_full, replacing the bag with a single model in a similar manner to use_child_oof=True, but supported for any model rather than only models that have use_child_oof=True supported.
• Fixed a bug where ParallelBagging ignored save_folds (SequentialBagging and ParallelBagging now produce the same result instead of different results when save_folds=False, which is relevant to reduce disk usage).
• Removed a redundant predict call on all of the data during refit_full via the new _skip_oof argument. This should drastically speed up refit_full for models that are slow in inference, and avoids needlessly storing unused OOF predictions in the model artifact, saving disk space.

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-3148-33e5307 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-3148/33e5307/index.html

[yinweisu]

LGTM!