This fixes an issue with `cosmic-panel` where, when a workspace is moved
back to an output after a monitor is disconnected and reconnected, the
panel doesn't hide because `cosmic-panel` thinks no toplevel is open on
that monitor.
After some testing, it seems `output_enter` isn't being sent here. In
particular, the `output_leave` call happens before the client binds the
`wl_output`, so there is no `wl_output` to send in an event yet.
This is addressed by keeping track of a set of `wl_output`s that we have
sent the event to. So if an output is bound, `refresh` can add it to
this list and send the event.
This is not needed for workspaces (though it could be done similarly)
since the handle objects are created by server events. So no race should
occur as long as the workspaces global is bound before the toplevel info
one.
Cosmic-workspaces was having an issue when the compositor sent `ready`,
`failed`, then `stopped`. This could be worked around client-side, but
presumably the compositor should never send more than one
`failed`/`ready` for a particular frame.
It seems cleaner to have `FrameInner::fail` as the only place that sends
`failed`. So any checks can be there. I believe the logic there should
be appropriate for every call.
Before this, `/proc/$(pidof cosmic-comp)/maps` quickly expands in length
when cosmic-workspaces is opened and closed a bunch of time, preventing
the GPU memory from being freed. Which on my Intel system can lead to
OOM eventually.
There may be other leaks to deal with, but `maps` no longer shows this
issue.
For the `drm` protocol, we can't construct an `ImportNotifier`, since
that is specifically based around `ZwpLinuxBufferParamsv1`. So we need a
new method for importing with that protocol.
This could be improved, but should match current behavior.
