Issue #, if available:

#3896
#3923

Description of changes:

Previously, seed_everything was called once at the beginning of TimeSeriesPredictor fit and predict methods which led to two undesirable side effects:

• if any of the models reset the seeds in their internal logic, they could effectively cancel the random_seed provided by the user. This was done, e.g., in RecursiveTabular's NN_TORCH affecting high_quality presets.
• The fitting order of models could affect the outcomes of individual models. i.e., the random state at the beginning of each model depended on previous models.

With this change, we propagate the random seed to the learner and trainer both at training and prediction time. We set the seed again, per model, when fitting single models or predicting with single models.

Also fixes a minor issue in predictor unit tests.

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