[ENH] set seed in test_score

[XinyuWuu]

This PR add seed setting in test_score in TestAllForecasters.

The DeepAR model in PR #6551 is giving different output every call to predict which causes problem with test_score.

As suggested by @benHeid, maybe we can pass some parameters to the predict function in pytorch-forecasting model to control the sampling behavior, howerver I can't find such parameters. Luckily after setting seed for some fundamental packages before every call to predict, DeepAR gives the same output every time.

As mentioned by @fkiraly, more than one deep learning models are facing the same problem. Setting seed in test_score for some fundamental packages could be helpful for models using sampling or other non-deterministic methods to generate prediction.

[XinyuWuu]

As discussed in the tech session, we might need to reset the seed back to original seed. We can do it for numpy, random and torch but not tensorflow.

[fkiraly]

To quickly summarize my feedback from the tech session:

• we need to make sure that setting of seeds is not global and affects other tests or estimators
• numpy.seed is afaik deprecated, generators should be used instead. This works via random_state for model that support it. Most deep learning models do not.
• related to the above but not this PR directly, most tensorflow based models fail tests that also assume we can derandomize, e.g., test_multiprocessing_idempotent, hence this test is skipped for them in tests._config. If we can find a way to set the seed on a per-estimator-instance basis for tf or torch, that would be great. See [ENH] randomization/derandomization tag and conditional test logic #6274