loads has no way to tell if it should provide text or bytes to the
serializer; bytes are more reasonnable for a serialized representation,
and are the only option for pickled data.
dumps can perform conversions on the value it receives from the
serializer; but for consistency it seems better to require bytes too.
The current code would cause an exception when loading pickled session
data. See next commit.
Also fixed a bug when checking for compressed data.