Fixes #198

Co-authored-by: Ben Kurtovic <ben.kurtovic@gmail.com>

Fixes #221.

Also removed the max cycles stop-gap, allowing much more complex pages
to be parsed quickly without losing nodes at the end

Also fixes #65, fixes #102, fixes #165, fixes #183
Also fixes #81 (Rafael Nadal parsing bug)
Also fixes #53, fixes #58, fixes #88, fixes #152 (duplicate issues)

Tests were not correctly testing the situations without a table close.
Fixed tests and then fixed tokenizers for failing tests. Also refactored
pytokenizer to more closely match the ctokenizer by only holding the
`_parse` methods in the try blocks and no other code.

Python 3.4 compiles C extensions with the
`-Werror=declaration-after-statement` flag that enforces C90 more
strictly than previous versions. Move all statements after declarations
to make sure this extension builds on 3.4.

Table tags no longer self-closing. Rows and cells now contain their
contents. Also refactored out an `emit_table_tag` method.
Note: this will require changes to the Tag node and possibly the builder,
those changes will be in the next commit.

CTokenizer is completely implemented in this commit - it didn't
make much sense to me to split it up. All tests passing, memory test
shows no leaks on Linux.

For the C tokenizer, include `<stdint.h>` and use `uint64_t` instead
of `int` for context. Changes to tables mean that context can be
larger than 32 bits, and it is possible for `int` to only have 16
bits anyways (though this is very unlikely).

* Removed unittest2 dependency on Python >2.6.
* Moved discover_tests.py into tests/.
* tokenizer.c: Fixed errors noted by -Wshorten-64-to-32.