Changed dynamic lexer behavior to only match terminals to their maximum length (i.e. greedy match), emulating the standard lexer.

The original dynamic lexer behavior, that attempts to match all appearances of a terminal, has been moved to the "dynamic_complete" lexer.

For example, when applying a terminal "a"+ to the text "aaa":

- dynamic: ["aaa"]
- dynamic_complete: ["a", "aa", "aaa"]

* All exceptions are now under exceptions.py
* UnexpectedInput is now superclass of UnexpectedToken and UnexpectedCharacters,
  all of which support the get_context() and match_examples() methods.

Nearley tool still needs fixing

Anonymous tokens would become visible if they had the same value as named tokens.
That's because they are merged for the lexer. But after this change, the rules for
visibility are based on their use in the rule, and not their name or identity.

* Removed the scanless parsing feature (dynamic lexing is king)
* Default LALR lexer is now contextual

(inspired by issue #116)

It emulates antlr's visitor behavior for a dynamic evaluation order of subtrees

I found this while porting Token to C, essentially the value and pos_in_stream members of Token were swapped in ``__reduce__``, which means running ``pickle.loads`` and ``pickle.dumps`` would result in unpickled tokens whose value was the original's position in stream, and vice versa. In my C extension this caused a TypeError exception, but the behavior will have to be corrected in both.

Added __slots__ to RulePtr and Token, resulting in significantly lower memory consumption.

As suggested by @drslump.