Spawn async tasks that need to outlive the calling process.

Used by the admin LiveViews to fire HTTP notifications after a DB commit (status change, delete, etc.) — the DB transaction has already succeeded, so the remote site should learn about it even if the admin closes the tab. Task.start_link/1 is wrong for this case because it cancels the task when the LV dies, and a bare Task.start/1 runs with no supervision: failures aren't logged and tasks accumulate on a hung HTTP timeout.

notify_remote_async/1 routes through PhoenixKit.TaskSupervisor (a named Task.Supervisor started by core phoenix_kit's supervision tree) with restart: :temporary. When that supervisor is not running (a stripped-down test environment; core starts it in every host) the task is started linked to the caller instead, so it is still owned by a process and still runs: it then ends with its caller rather than outliving it, and the fallback is logged.