feat(api): add non-blocking Tracker close-and-drain completion receipt

[ThoriaDevelopment]

Summary

Implements #417: a non-blocking completion receipt that proves a Tracker has fully drained before the plugin removes the owning entity.

Tracker#closeAndDrain() closes the tracker and returns a CompletionStage<Void> that completes only when all tracker-originated work has ended:

1. no new tracker-originated source or pipeline work can begin;
2. scheduled recurrence has been cancelled;
3. any updater invocation already running has returned;
4. queued or deferred tracker work has either completed or been invalidated;
5. tracker-originated callbacks and tasks that could access the source have completed or been invalidated;
6. pipeline despawn and final close work has completed; and
7. after completion, no future tracker-originated access to the source occurs.

It never blocks and is safe to call from the main server thread:

tracker.closeAndDrain()
    .thenRunAsync(() -> {
        // safe: the tracker can no longer access the source
        entity.remove();
    });

Impacted modules

• api (kr.toxicity.model.api.tracker.Tracker, CloseReason.DRAIN)
• api test source set (new: TrackerCloseAndDrainTest + JUnit test dependencies in api/build.gradle.kts)

How it works

• A quiescence gate inside tick(): once draining is requested, the next completed updater invocation (including one already in flight) runs any still-queued tasks and completes the receipt exactly once.
• Completion attempts are ordered by the tracker monitor and the ticking flag is set before the quiescence check, so a close racing a running updater can never complete the receipt while the updater is still inside its handler (Dekker-style ordering).
• close(CloseReason.DRAIN) cancels the scheduler without interrupting the in-flight updater — interrupting could abort pipeline work mid-flight before the receipt completes. Legacy reasons (REMOVE, DESPAWN, PLUGIN_DISABLE) keep their previous interrupt semantics.
• task(...) rejects new work from the moment draining begins; tasks queued before the drain are flushed before completion or invalidated.
• start() now refuses to reschedule a closed tracker (a late spawnPacketHandler callback could previously restart the scheduler after close).
• pipeline.despawn() runs before the receipt completes, satisfying postcondition 6; the constructor's late skin-context callback is invalidated once the tracker closes, satisfying postcondition 5/7.
• If a queued task or close handler throws, the receipt completes exceptionally instead of hanging; repeated/reentrant calls return the same stage.

Compatibility

• Purely additive public API (closeAndDrain(), @since 3.4.2, matching the repo's @since versioning convention). No existing signatures changed.
• close() and shutdown() keep the same interrupt/despawn semantics for all pre-existing reasons. One deliberate behavior change: a throwing close handler is now logged and still fully tears the tracker down (previously teardown was skipped and the exception propagated).

Tests / checks run

• TrackerCloseAndDrainTest (7 tests, all green): immediate completion when idle; queued-task invalidation after drain; queued work runs before completion; latch-blocked updater defers the receipt until it returns; reentrant/idempotent closes; exceptional completion for failing queued tasks and close handlers.
• :bettermodel-api:check and :bettermodel-api:checkLicenses — pass.

---

This contribution was prepared with AI assistance (Codebuff) and reviewed by a human before submission.