Hide the entry before detaching it in _release_ui (shell crash on GNOME 50)

[daniel-g-carrasco]

Launching an app from the overlay on GNOME 50 sometimes aborts the whole shell (on Wayland that means losing the session, and GNOME's crash protection then disables all extensions):

Clutter:ERROR:../clutter/clutter/clutter-actor.c:1989:clutter_actor_real_unrealize:
  assertion failed: (!clutter_actor_is_mapped (self))
== Stack trace for context ==
#0  extension.js:755 (_release_ui)
#1  extension.js:495 (hide)
#2  extension.js:1012

Caught it with a coredump on Fedora 44 / gnome-shell 50.2 / Wayland. _release_ui() calls remove_child() on the entry while the overlay is still mapped, and clutter 18's stricter unrealize assertions abort on that.

Hiding the entry first unmaps it so the detach is safe. There's no visible difference since the whole overlay is fading out at that point anyway.

This is probably what #82 and #133 have been hitting too.

[oameye]

oeoeoeo exactly the fix I am looking for!

[daniel-g-carrasco]

Pushed a follow-up commit: a re-entrancy guard on hide().

While validating this fix on GNOME 50 I hit a second instance of the same abort, reached from the keyboard accelerator (Super+Space) rather than a click:

clutter_actor_set_mapped: assertion '!CLUTTER_ACTOR_IN_MAP_UNMAP' failed
Clutter:ERROR clutter_actor_real_unrealize: assertion (!clutter_actor_is_mapped (self))

Root cause (coredump JS stack): the this._entry.hide() this PR adds flips the stage key-focus → _onKeyFocusChanged → this.hide() again → re-enters _release_ui() while the first teardown is still mid-unmap. The guard (this._hiding) makes the recursive hide() a no-op so teardown runs exactly once. With both changes, click and keyboard paths are stable here.

[daniel-g-carrasco]

Folded the full root-cause writeup + the remaining crash sites into #166 — this PR is the _release_ui hide-before-detach + hide() re-entrancy part of it.