device->authr cannot be reset to NULL in initiator case as it is needed
by device_is_authenticating.

So instead of reseting device->authr just reset both device->authr->cb
and device->authr->agent to NULL, this fix the regression since the original
problem was device->authr->agent pointing to freed memory.

During general bonding or dedicated bonding acceptor
device->authr->agent will point to the adapter agent. However, there's
no direct way to access device->authr if the agent exists while
authentication is ongoing. With the current code this would lead into
accessing free'd memory since device->authr->agent becomes an invalid
pointer.

One way to fix this would be to add reference counting to the agent
objects, but that would be quite intrusive on the way that the code
expects agent_free to behave at the moment and handles agent
re-registration checks. Another option would be to iterate through all
device objects to find if any of them have a reference to the adapter
agent, however this would be quite a waste of CPU cycles (and pretty
ugly imho).

This patch takes neither of these two possibilities but comes with the
realization that maintaining device->authr is completely unnecessary
after the authentication callbacks have been called (currently
device->authr gets reset only when the full pairing process completes).
Since agent_free takes care of the bluetoothd internal callbacks in the
case that the agent exited unexpectedly we simply need to clean up
device->authr in these callbacks, which is what this patch does.

This is needed until everything has a proper btd_ symbol prefix.

It seems the play and pause key presses also lack the release event
sometimes so also add the necessary quirk to them.