Get rid of the `random_state` argument in `get_pipeline`

Currently, the best_pipeline property in AutoMLSearch will always return a pipeline with random_state=0 because that’s the default value for random_state in the get_pipeline method.

I think this is confusing because if I create an AutoMLSearch object with random_state=5, for example, I expect all of the pipelines created and returned by AutoMLSearch to have random_state=5.

I think we should remove the random_state argument from get_pipeline and instead make sure that the random state is set to whatever it was when the pipeline was fit.

[bchen1116]

Through discussion with @freddyaboulton, we decided it’s fine to have the pipeline random_state be a np.RandomState object, but we want to make sure it’s seeded with the random_state int that a user passes in, OR with the np.RandomState object that a user passes in, depending on the input.

This means that if a user passes in random_state=4 to AutoMLSearch, we create the pipeline random_state with ‘4’ as the seed, and if a user passes in random_state=np.RandomState(), we pass that to create the pipeline instead.

[bchen1116]

@freddyaboulton I think the largest issue is that in our pipeline_base, we set the random_state to be an np.RandomState object, as seen here

If we remove the np.RandomState object, then we can preserve the random state that the user could pass in, but we would also need to change the default random_state behavior for component_base as well. This issue does seem to be tied with a discussion about support for np.RandomState. What do you think?