fix(windows-lib): release thumbnails after rendered frame

[sim590]

Problem

Live window thumbnails leave a significant amount of GPU memory allocated
after the overview or switch surface is closed. Reopening the surface can
also leave stale GLib callbacks behind, causing the next invocation to fail
or crash.

Root cause

Thumbnail widgets keep references to GDK textures backed by imported DMA-BUF
images. Hiding the layer surface before GTK renders an updated scene prevents
GSK from observing that those images are no longer used. The images are then
kept by the GSK GPU cache until its normal garbage-collection timeout.

Fix

Clear the thumbnail widgets before hiding the surface and wait for the
GdkFrameClock after-paint signal. This ensures that a frame without the
thumbnail textures has been rendered before the surface is hidden.

The overview waits for all monitor windows to complete this step before
performing a pending client or workspace switch. Frame handlers are
disconnected safely when a surface is reopened.

Refresh callbacks use cancellation tokens instead of removing potentially
expired GLib sources, avoiding crashes during rapid close and reopen cycles.

DMA-BUF lifetime

Keep a duplicated file descriptor alive for each GDK DMA-BUF texture instead
of relying on a bounded list of retired GBM buffer objects. This makes the
DMA-BUF lifetime follow the texture that imported it and prevents closed file
descriptors from being reused after buffer reallocation.

This is an independent lifetime-safety fix for resize and reallocation
scenarios; it is not required for the normal VRAM cleanup path described
above.

[sim590]

@H3rmt: I noticed that VRAM was still allocated after spawning and hiding the live thumbnails.

I deduced that this was memory allocated in the GSK cache that needed the proper signal to get freed. You'll notice a delay before it clears. This is due to the internal cache clearing timeout which defaults to 15 seconds. You can see no delay if you set GSK_CACHE_TIMEOUT=0.

All of this seemed to happen because window.set_visible(false); is called too early. In fact, if you do the following edit to my branch:

diff --git a/crates/windows-lib/src/switch/root.rs b/crates/windows-lib/src/switch/root.rs
index 10fa474..46b3bac 100644
--- a/crates/windows-lib/src/switch/root.rs
+++ b/crates/windows-lib/src/switch/root.rs
@@ -582,6 +582,7 @@ impl SwitchRoot {
                 });
                 *self.after_paint_handler.borrow_mut() = Some(id);
                 self.window.queue_draw();
+                self.window.set_visible(false);
             } else {
                 // No frame clock available: hide directly instead of waiting.
                 self.window.set_visible(false);

You'll notice that the VRAM is not freed at all anymore. This is because the window is hidden immediately after requesting the redraw, before GTK has a chance to process and render the queued frame.

[sim590]

If you prefer, I could split the two fixes in two PRs. I just felt that the DMA-BUF lifetime fix could fit in the same patch even though it's not required.