Implement Joop Leo's optimisations for right recursion

We can replace Earley Columns with basic python sets
for improved performance and simplicity.

When using the same parser repeatedly for small parsers we incur
significant overhead by recreating the ForestVisitor each parser.
We can cache the Forest walker and re-use it by making it stateless.
Also, we can use slots for all of the Forest Walkers to reduce
construction delay and function call overhead.

Fix typo in docs

Instructions how to get Shift/Reduce messages printed (Issue #258)

Method iter_subtrees_topdown added

types of ambiguity.
 - Rewritten along the lines of Elizabeth Scott's parser.
https://www.sciencedirect.com/science/article/pii/S1571066108001497

 - Implement SPPF trees per Elizabeth Scott and Bram van der Sanden's work.
http://www.bramvandersanden.com/post/2014/06/shared-packed-parse-forest/

Allow to passing "rankdir" to lark/tree.py:pydot__tree_to_png

lark/tree.py

shaping. For example, passing "TB" instead of "LR" makes the tree
to be draw vertically instead of horizontally.

https://stackoverflow.com/questions/29003465/pydot-graphviz-how-to-order-horizontally-nodes-in-a-cluster-while-the-rest-of-t

docs: clarify end_column value

I interpreted "the column where the token ends" to mean "a pointer
to the last character in the token", which would be the same as
the start column for a single-character token.

However, that's not what lark actually reports. Reword to clarify
this.

https://github.com/lark-parser/lark/issues/240

Added missing binary files to .gitignore