In Python 2.x, if you write to stdout and stdout is piped
(for example: `python -m markdown foo.txt | less`), then
`sys.stdout.encoding` is `None` and an error is rasied.
Commit 1132f9e20cd7a5d6be809651f1034c44c32dbc0e was an attempt to
fix this, and it works in Python 2.x.

However, it does not work in Python 3.x, which does not exhibit this problem.
In fact, that fix actually breaks things in Python 3 whether the output
is piped or not. Additionaly, in Python 2.x, the fix is not needed if the
output is not being piped.

As we do not have a version specific issue, but an issue with
`sys.stdout.encoding`, we check for that to determine which way to go.
This way, the "right thing" *should* happen every time.

Specificaly, `self.output_format` is defined and contains a string of the
output format used on the instance. This is more useful that an instance
of the searializer when determining alternate behavior elsewhere in the parser.
For example, see Issue #129.

Also cleaned up the error when an invalid format is provided. We now re-raise
the original error (with a custom message) rather than raising a new error.

In this case text should be encoded into the output encoding explicitly, because
sys.stdout.encoding is None, when piping data.

Note that you can still explicitly set 'enable_attributes' and that
value will be honored regardless of 'safe_mode'. However if 'safe_mode'
is on and 'enable_attributes' is not explicitly set, then
'enable_attributes' defaults to False.