Cloning and serialization improvements

[engintoklu]

The main purpose of this pull request is to increase the correctness when cloning and/or pickling instances of Problem, Solution, SolutionBatch, ObjectArray, and ImmutableContainer. When cloning and/or pickling, two new behaviors are introduced:
(i) PyTorch tensors are cut off from their original storages (to avoid small slices from carrying their original storages with themselves); and
(ii) self references within these mentioned objects are specially handled, and therefore, many unwanted RecursionError scenarios which could previously happen are now avoided.

New function: deep_clone(...)

This new function creates a deep copy of a Python container (which can contain other containers or PyTorch tensors or numpy arrays). While traversing through the container, any PyTorch tensors will be copied using the .clone() method, which means, in the newly created clone, the PyTorch tensors will be independent copies without extra storage.

New base class: Clonable

Any subclass of Clonable is expected to override the method _get_cloned_state(...). This method is expected to return the state of the object in a dictionary. The Clonable base class provides definitions for the methods __copy__(...) and __deepcopy__(...) where the dictionary made by _get_cloned_state(...) is used to create the clone.

New base class: Serializable

This is a subclass of Clonable, which is meant to be used as a base class. In addition to the method definitions made by Clonable (__copy__ and __deepcopy__), this base class also defines __getstate__ so that the method _get_cloned_state(...) also affects how the inheriting class is pickled.

Classes inheriting from Serializable

The following classes now inherit from Serializable:

• Problem
• Solution
• SolutionBatch
• Distribution

Each of these mentioned classes have their own definitions for _get_cloned_state(...), and use deep_clone(...) for preparing their state dictionaries. Therefore, at the moment of cloning and pickling, the stored tensors are subject to the .clone() method and they are cut off from their original storages. With this feature, when a relatively small slice of a large SolutionBatch is cloned or pickled, the serialized slice will not contain the entire storage of the original batch.

New base class: ReadOnlyClonable

This is like Clonable, but requires the inheriting class to also define a method called _get_mutable_clone() so that the object knows how its mutable counterpart can be constructed. ImmutableContainer objects are now ReadOnlyClonable.

[NaturalGradient]

Excluding the gym compatibility issues described in #22 , every existing example appears to work as expected. I have also reviewed the code and have found no notable issues except for those already discussed and clarified by @engintoklu .