I spelled it "fourty", which tricks the lmtool into making it much
harder to recognize the word.  Now it's spelled "forty", which makes it
much easier to understand.

Added some blank lines and wrapped some long lines to be closer to PEP 8
compliance.

Improved docstrings in NumberParser class.  Added docstrings for
undocumented code and removed a TODO message that no longer applies.

That's "Kaylee Voice Command" if you didn't figure it out.  I think
everything still works properly, but I'll do more testing later to
verify.

Its control flow was confusing before; now it's much more
straightforward.  We make a string representing classes of words, split
that by a regular expression for number words, then parse each number
and build up our return string and list.  It works just as well as the
previous method, is a bit shorter, and I feel that it's clearer as well.

It wasn't actually guaranteed, it turns out.  I was iterating over
dictionary keys, which is done in arbitrary order.  The result was that
in different executions of the program, the corpus was generated
differently, so the hashes differed, and the language had to be updated.
Sorting the keys before adding them to the list of number-words fixed
the problem.

See commands.tmp for an example.  It's pretty neat, but it could still
use some work.  I thought of a really clever way to parse numbers, better
than the one I came up with last night, but since I have a working
implementation now I figure I'd better commit it.

We have a new bug which causes the dictionary to be updated every time the
program starts.  I hope I didn't force that to happen last night or
something, but I have a vague feeling I did.