fix production mode and scheduler race

[cebtenzzre]

Problem

Commit 2239e58 ("Merge branch 'context-managers'") from 4.5.12 caused regressions in app initialization.

• Apps would no longer load with production_mode: true because the async task races with initialization, and check_app_updates() is never called again.
• Even in debug mode, app initialization became non-deterministic.

The system would log "App initialization complete" but apps weren't actually loaded yet (or ever, in production mode).

Root Cause

The old app_management.start() method used loop.create_task() to initialize apps asynchronously:

def start(self):
    self.AD.loop.create_task(self.init_admin_entities())
    task = self.AD.loop.create_task(
        self.check_app_updates(mode=UpdateMode.INIT),
        name="check_app_updates",
    )
    task.add_done_callback(lambda _: ...)

This created a race where apps might not be loaded before the system continued. In production mode, where check_app_updates() is skipped in the utility loop, apps would never load if the task hadn't completed yet.

Solution

app_management.start() is replaced with a synchronous version. Scheduler initialization and app loading are moved back to the end of Utility._init_loop.

Two regression tests are added to verify that production mode is once again functional.

Fixes #2442
Fixes #2467

[mmattel]

@jsl12 @acockburn note, if this gets merged, a new AD version should be created and published...

[jsl12]

Looks good to me