Right now MLEM fails if you install it with pip install mlem, cause numpy is not in install_requires.

Hey @aguschin @mike0sv , I just noticed this issue I introduced in my previous PR, sorry about that 🙏🏻 I thought through more deeply, and I think this merging of results should be dependent on different model types. What do y'all think about introducing a method on ModelType class, eg. merge_results, that take an array of a certain type and combine the results accordingly based on their types (e.g numpy array in the case of sklearn predict, predict_proba methods). In this case, commands.py will not need to know model-specific details.

Actually this should be a method of DatasetType, since the same model can output different types, but the idea is the same