Script completions no longer stall portal shutdown

The very first time a portal process ran a script cell or offered code completions in a script-flavored editor, it had to build a shared list of reference assemblies — a one-time cost that is normally quick, but occasionally ran long under memory pressure. That build used to run inline while holding a slot in the pool a shutdown needs to hand back cleanly, so on the rare occasion it was slow, a restart or redeploy could sit waiting on it for up to 30 seconds instead of completing promptly.

The build now runs independently of any one request: a request waits for it but no longer blocks a shutdown from proceeding, even on that first, slower call. Restarts and redeploys are consistently quick regardless of when they land relative to the first script or completion request.