Summary

Fixes #4911 — VowpalWabbitThreadedLearning was nearly impossible to use because SaveModel() and PerformanceStatistics had to be called before Complete(), which is the opposite of what users expect.

Root cause: Complete() immediately called syncActions.CompleteAdding(), closing the queue before the completion continuation (which actually drains it) had a chance to run. Any SaveModel() call after Complete() threw InvalidOperationException.

Changes:

• Allow SaveModel/PerformanceStatistics after completion — detect post-completion state and save directly on the root VW instance. Uses TryAdd on the sync action list as a fallback for the race window between Complete() and the continuation.
• Move CompleteAdding into the root completion continuation — uses a new atomic CompleteAndRemoveAll() on ConcurrentList<T>, so sync actions can be enqueued between calling Complete() and the continuation executing.
• Add Flush() method — forces an AllReduce synchronization and drains pending sync actions on demand, without requiring the example count to hit a multiple of ExampleCountPerRun.

All changes are backward-compatible — existing code that calls SaveModel before Complete() continues to work.

The natural pattern now works:

await model.Complete();
await model.SaveModel(path);   // no longer throws
var stats = await model.PerformanceStatistics;  // also works

Mid-training saves also work without count alignment:

var saveTask = model.SaveModel(path);
await model.Flush();   // forces sync now
await saveTask;
// continue learning...

Test plan

• TestSaveModelAfterComplete — save after Complete() produces a valid model file
• TestPerformanceStatisticsAfterComplete — stats accessible after Complete()
• TestFlush — Flush() triggers sync and save mid-training, learning continues after
• TestSaveModelBeforeComplete — original pattern (save before complete) still works
• Existing TestAllReduce continues to pass (CI)