In the case that the AVDTP session references weren't fully dropped when a
disconnection happens, in later reconnections the disconnect_timeout
function cound have a reference count higher than 1. These bugs naturally
need to fixed (and there's a proper error() call to log them) but we can
still handle the situation semi-cleanly instead of asserting and aborting
since the AVDTP session struct will just silently hang around and get
reused at the next connection attempt.

The regression was introduced by audio_device_request_authorization which
cause the authorization request to be added the list no matter what
btd_request_authorization returns.

The error being returned for hang up currently is
org.bluez.Error.ConnectionAttemptFailed which is not accurate since the
connection itself was already established.

This should fix it by translating the hang up to
org.bluez.Error.AuthenticationCanceled.

Like with the A2DP signalling channel this makes sure no process reads or
writes the SCO channel after bluetoothd wants it to be closed.

At the moment of writing this patch the kernel doesn't seem to handle this
right so a fix is needed on that side too for the intended effect to take
place.