feat(oci): GPU selection per host and per image, one notification per update, and App tab for stack containers

- Immich asks what runs its video and its recognition in one menu, in both
  modes, and gives the GPU to the server and to Machine learning; AMD uses ROCm
- Frigate, Ollama, llama.cpp, Faster Whisper and Piper take the image built
  for the chosen GPU
- The acceleration menu offers only what the host can run
- An update or a recreation sends one notification with its result instead of
  the stop, backup and start of each container
- A private bridge with nothing connected is not reported as down
- Secondary containers of a stack appear in the App tab with their version and
  logo; Secure Gateway shows the same update state in both views
- A mistyped value in the wizard asks the same question again

oci/remote/oci_native_stack.py

@@ -40,6 +40,9 @@ def begin(root, primary, template, deployment, members, adapter):
                 'mounts': []}
         if Path(adapter).name == 'install_immich_stack.sh' and name == 'machine-learning':
             plan['machine_learning'] = copy.deepcopy(deployment.get('machine_learning', {'acceleration': 'cpu'}))
+        if Path(adapter).name == 'install_immich_stack.sh' and name == 'server':
+            # An update must know that the server transcodes with NVIDIA.
+            plan['video_transcoding'] = copy.deepcopy(deployment.get('video_transcoding', {'acceleration': 'cpu'}))
         if Path(adapter).name in oci_stack_replay.FILES:
             plan['replay_profile'] = {'adapter': Path(adapter).name, 'role': name}
             if not oci_stack_replay.FILES[Path(adapter).name][name]: