Frozendict

* speedup to hash computation
* subclassing frozendict now is safer
* now an exception type and message that is raised by immutability is aligned to the stdlib (tuple, frozenset)

Turned back to the `dict` API:

* removed `is_frozendict` and `initialized` attributes. Added `_initialized`.
* removed `__add__()` and `__sub__()` methods (wait for [PEP 584](https://www.python.org/dev/peps/pep-0584/))

Many, Many changes under and over the hood:

* now `copy()` and `frozendict(another_frozendict)` returns the same instance, like all the other CPython immutables
* fixed `repr()`: previously the result string was cached. Since now `frozendict` can contains also mutables, this is wrong (and maybe it is wrong also when all the values are immutable...)
* `a_frozendict - iterable` now supports also text-like (`str`, `bytes`...) and generator-like iterables.
* Removed the not-so-useful `frozendictbase` class.
* Many micro-improvments to speed
* Now attribute `initialized` is public

- Now you can add also unhashable values to a `frozendict`. If you try to get the hash of such a frozendict, a TypeError will be raised. This is the same behaviour of `tuple`, `immutables.Map` and the slezica's `frozendict`
- The empty `frozendict()` is now a singleton, like `()` and `frozenset()`.
So `(frozendict() is frozendict([])) == True`.

Code and build cleaning.

Speed improvements to constructor.