Despite the old comments, re-initing the msg_t leaks a refcount to
metadata in some situations.

v1_decoder looks like it isn't tested any more, but it seems like a good
idea to fix it because it has the exact same piece of buggy code
v2_decoder does.