[#106] fix GeneratorList with Python 3 (add __next__)

Credit for this fix goes to @tosher:
https://github.com/mwclient/mwclient/issues#issuecomment-154751657
I just figured out exactly what changes his comment suggested,
and poked around a bit to figure out what was going on and why
they're needed.

`GeneratorList` (and things deriving from it, e.g. `Category`)
was broken for iteration purposes in Python 3. In Python 3
the iterator protocol requires a `__next__` method, not a `next`
method as in Python 2. 8d0650cd renamed `List`'s `next` method
to `__next__` and added a small wrapper `next` method to account
for this, but did not make the same change for `GeneratorList`.
So when you iterate over a `GeneratorList` with Python 3, it
winds up calling `List.__next__` (since `GeneratorList` inherits
from `List`), not `GeneratorList.next`. So we don't hit the bit
that turns the results into mwclient page objects, and just get
the raw dicts from the `List` method.

We also have to change a couple of direct calls to `List.next`
to call `List.__next__` instead. If we don't do this, then when
the call originates from a child class, things get messed up,
because `List.next` calls `self.__next__` - and that will wind
up calling the child class' `__next__`, not `List`'s `__next__`.

Confused yet? :)

Fix two typos

- Removing the need to use a double for loop to consume the the result
  by returning a list of revisions rather than a list of pages.
- Parsing the timestamps using the standard parse_timestamp method
- Adding tests

(#84)

Test in Python 3.5 too, now that it is out

Cache page text until next edit operation

Store the results of page.text() operations in a simple cache
dict. This avoids unnecessary remote roundtrips. Cache is
cleared on each successful page.save() operation. cache
argument can be set to 'False' to disable use of the cache.

Remove unused local vars

This attribute goes back to an older design, where
the 'edit page' workflow was different: the edit() method did
part of the work of defining what text you meant to edit. In
the current workflow, text() - which replaced the deprecated
edit() - is only for retrieving text, and save() is supposed
to do all the work of editing the page.

The 'section' attribute is therefore obsolete and dangerous.

- Remove :ref:s from the headers, since they made the latex build crash
- Add logo

Remaining issue:

- Can't get headings levels right on both the html and the latex builds
  at the same time. More info in #40

While the 'backlinks' method has a 'redirect' option,
there is no such option for the 'embeddedin' method.

See 90b87566 for context

Also change link to the slightly user-friendly Phabricator (the url looks nicer without the escaped slash, too)

Removing unsued import

From client.py: time, random, sys, weakref
From sleep.py: random

Merging changes from mwclient master

Since CirrusSearch (and Lucene?) doesn't support 'title'.

Since default values are evaluated only once, mutable objects shouldn't be used.

I've no idea when this would actually be needed, but at least it should be fixed now.. It broke during the move to use the requests library.

log.warn has been deprecated