This adds a new `randomize_seed` for regression and classification. When not explicitly set, the seed is randomly generated. One can reuse the seed in a similar job in order to ensure the same docs are picked for training.
