revised patch for gmime init, with test.

[David Bremner]

It turns out that our existing (trivial) python test is enough to
catch this bug, but the corpus needs to be augmented.  This
augmentation is a bit intrusive so I'm thinking of cherry-picking only
the actual fix to the release branch.

Unfortunately the test message is 8 bit, so it may be encoded in some
inconvenient way for patch application. The message is attached to an
earlier message in the thread if you want to double check.

I also wondered about putting g_type_init inside the (!initialized)
test, but decided against it on the grounds of minimality.

I think we want to in the medium term factor out all of the
initialization code into one (probably private) function; we can clean
things up a bit more then.