Return model predictions as dataclasses instead of pydantic models

[kmaziarz]

The Prediction and PredictionList objects were made based on pydantic, anticipating this will make it easier to e.g. expose a model through an API using a framework such as FastAPI. However, this meant introducing pydantic into the core of the library in order to facilitate a feature that most users may not care about. On the other hand, FastAPI can deal with non-pydantic classes too (pydantic is only needed for some extras like validation). Thus, once we want to integrate the server code properly, we should do it in such a way that pydantic (and further dependencies like fastapi) remain optional extras rather than part of the core of the library.

For now, in this PR I make the core prediction objects plain dataclasses, and drop pydantic from dependencies. This required a few changes (e.g. dropping special handling of prediction classes in dictify, using default_factory for the dict-valued fields), as well as fixing some of the tests, which did not provide all fields needed to construct the prediction objects (which for some reason is still accepted by pydantic).

[AustinT]

LGTM!