ci(smoke): fix state tier hang from bare wait on a lingering daemon

The state tier started actually running mpd in 74f0d33, which exposed a
deadlock in teardown: state.sh launches `mpd --no-daemon &` and never kills
it during the tier, and its cleanup() called smoke::stop *before* killing mpd.
smoke::stop ended in a bare `wait`, which reaps *every* background job of the
shell -- including the still-running mpd -- so it blocked until GitHub's 6h
job timeout. continue-on-error doesn't help: it catches failures, not hangs.

- lib.sh: smoke::stop now waits only on the PIDs it owns (waybar, compositor),
  so an unrelated daemon left running by a tier can't deadlock teardown.
- state.sh: cleanup() tears down mpd/pulseaudio before smoke::stop.
- smoke.yml: add timeout-minutes: 20 so a future hang fails fast instead of
  burning the default 6h runner budget.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

test/smoke/lib.sh

@@ -194,5 +194,12 @@ smoke::assert_clean_exit() {
 smoke::stop() {
     kill "$WAYBAR_PID" 2>/dev/null || true
     swaymsg -q exit 2>/dev/null || kill "$COMP_PID" 2>/dev/null || true
-    wait 2>/dev/null || true
+    # Wait only on the processes we own. A bare `wait` reaps *every* background
+    # job of the caller's shell, so a tier that leaves an unrelated daemon
+    # running (state.sh's mpd, killed later in its own cleanup) would deadlock
+    # here until GitHub's 6h job timeout.
+    local p
+    for p in "$WAYBAR_PID" "$COMP_PID"; do
+        [ -n "$p" ] && wait "$p" 2>/dev/null || true
+    done
 }