recommender improvements: base random_recommendation method

[lacava]

it would be nice to have a base method that generates a random ml + p recommendation with options:

• with or without uniformly sampling the ML options, and
• filtering based on trained_dataset_models

now that trained_dataset_models is centralized this should be doable.

this method could be used by recommenders if the n_recs is below the desired number due to repeat recommendations.