feat(oci): recover applications after a Proxmox reinstall, cluster records and AMD GPU profiles

The installation record of an OCI application travels with its container:
a copy inside the container and another in /etc/pve, written together
with the one kept on the host. A container restored on a newly installed
Proxmox, restored with another ID or moved to another node of a cluster
is recognised and registered again, with its private network, hookscript,
Rclone mount, host firewall rule and NVIDIA runtime. The Monitor offers
the same recovery from the Updates tab.

AMD GPUs are offered by generation. A GPU the ROCm image supports takes
the profile as it is; one of a supported family (Radeon 680M, 780M) is an
experimental option that asks for confirmation and is never proposed; an
older one is not offered. The GPU is checked with a real inference before
the installation accepts it. Recreate changes what runs recognition in an
installed Immich, between the CPU and a GPU of the host.

Updates:
- A failed update that is restored and checked removes its temporary
  container and the disks of the failed attempt.
- Every container volume is part of the backups, so Jellyfin, Plex and
  Hugo update with their default installation.
- An image published with a Docker-format manifest is recognised by its
  layers and build time and updates.
- The Proxmox notes of a multi-container application link to its LAN
  address.

oci/remote/oci_instance_transaction.py

@@ -764,9 +764,11 @@ def restore_description(vmid, state):
     run('pct', 'set', str(vmid), '--description', original_description(state))
 
 
-def release_stage(state):
+def release_stage(state, discard=False):
     """After a commit the holder CT only keeps its own rootfs: every parked
-    volume went back to the application. Anything still attached keeps it."""
+    volume went back to the application. After a verified recovery it holds
+    the disks of the failed attempt, which the restored backup replaced:
+    `discard` removes them with it. Otherwise anything still attached keeps it."""
     stage = state.get('stage')
     if not stage:
         return
@@ -774,11 +776,23 @@ def release_stage(state):
         config = owned(stage, 'proxmenux-transaction=' + state['id'])
     except (ValueError, RuntimeError, subprocess.CalledProcessError):
         return
-    if mounts(config) or any(re.fullmatch(r'unused[0-9]+', key) for key in parse_config(config)):
+    held = mounts(config) or any(re.fullmatch(r'unused[0-9]+', key) for key in parse_config(config))
+    if held and not discard:
         return
+    if discard:
+        stop(stage)
     run('pct', 'destroy', str(stage))
 
 
+def discard_stage(state):
+    """Remove the holder CT of an operation that ended in a verified recovery.
+    A failure here leaves it in place and does not undo the recovery."""
+    try:
+        release_stage(state, discard=True)
+    except (OSError, ValueError, RuntimeError, subprocess.SubprocessError) as error:
+        log(f'cleanup: {error}')
+
+
 def gib(size):
     return f'{size / 1024**3:.1f} GB'
 
@@ -1124,7 +1138,9 @@ def complete_recovery(root, journal, state):
     instances.write(instances.location(root, vmid), restored)
     checkpoint(journal, state, 'rolled-back')
     if show:
-        msg_ok(translate('Recovery completed. The displaced disks and the backup are kept; nothing was deleted automatically.'))
+        # A member of a stack is released when the whole stack is back.
+        discard_stage(state)
+        msg_ok(translate('Recovery completed. The disks of the failed attempt were removed; the backup is kept.'))
         if state.get('original_host_sources') or state.get('desired_host_sources'):
             msg_info2(translate('Shared host files are kept as they are; the backup does not restore their content.'))