Server salts change every 30 minutes after all, so keeping them
in the long-term storage session file doesn't make much sense.
Saving the layer doesn't make sense either, as it was only used
to know whether to init connection or not, but it should be done
always.
The reasoning behind this is that .connect() should not call
any request at all, it should only connect to the servers
although it currently still calls GetStateRequest.
There were some issues (#291, #360) where the auth_key was None
(possibly due to .connect() returning False), so this may
fix some of the cases where it returned False. This way we also
ensure that we always have an auth_key, or even if it "breaks"
(it's not the right key for the server anymore).
A few additional changes have been introduced to accommodate
this, such as moving InitConnection logic too or importing auths.
Every update that hadn't been acknowledged on the main connection
yet would be resent on any new connection. These new connections
are made temporary when invoking anything from any thread that's
not the main thread. It would also process all the updates, hence,
Telegram would be resending these not-acknowledged updates to the
temporary connection, and the updates would be processed again,
then the update handler would react to the duplicated updates over
and over.
To fix this, simply don't process updates on the temporary thread
at all. With this reasoning, if we don't acknowledge updates on
the temporary connections, Telegram will resend them on the main
connection, so we should not lose any.
This reverts commit 280a700655.
The reason for this is that it was causing a lot of files to
be downloaded corrupted for some reason. This should be
revisited to avoid creating a new connection for every chunk.
