[Multimodal] Fix seed

[yinweisu]

Issue #, if available:
#3219

Description of changes:

• The issue above can be reproduced by providing a fixed HPO search space, i.e. learning_rate = tune.choice([1]) and running multiple runs will give different results.
• The root cause should be because the current seed_everything call happens in fit while this seed setting is not affecting hpo trials. This behavior can be verified by printing torch seed inside _fit via torch.random.initial_seed(). Different runs will give different seed value.
• This PR moves the seed_everything call to _fit, which happens inside forked process created by ray and should fix the issue

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-3222-4da72e4 is done.
Docs are uploaded to http://autogluon-staging.s3-website-us-west-2.amazonaws.com/PR-3222/4da72e4/index.html

[zhiqiangdon]

LGTM. Thanks for the fix!