build-image: serialize host-wide with flock, clean up stale mounts/devices

Gitea runs each matrix entry as its own task, so build-image.yaml's
`max-parallel: 1` never serialized the two editions -- both builds started
at once on the fedora:host runner and corrupted each other's global
anaconda/livemedia-creator state (/mnt/sysroot, /mnt/sysimage, loop and
device-mapper devices), surfacing as bootctl failing with "Couldn't find
EFI system partition" even though the dish's EFI partition was fine.

build-image.sh now takes a host-wide flock around livemedia-creator, and
under the lock first clears any stale mount/dm/loop device a previously
failed build left behind. Verified on the runner: both editions launched
concurrently now run serially under the lock and build successfully.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

.gitea/workflows/build-image.yaml

@@ -42,8 +42,8 @@ jobs:
     runs-on: fedora
     strategy:
       fail-fast: false
-      # One image at a time: each build takes loop devices, RAM and tens of GiB
-      max-parallel: 1
+      # NB: Gitea runs every matrix entry as its own task, so `max-parallel`
+      # does not serialize them; build-image.sh takes a host-wide flock instead.
       matrix:
         include:
           - edition: phyllomeos