Fix deadlock in DrainAsync causing Redis scheduling test failure

[jeremydmiller]

Summary

• Fixes a deadlock in DurableReceiver.DrainAsync() and BufferedReceiver.DrainAsync() introduced by Latch receivers on ApplicationStopping for immediate graceful shutdown #2288
• The WaitForCompletionAsync call now only executes during actual shutdown (when Latch() was called separately via OnApplicationStopping), not during pipeline-triggered pauses

Root Cause

PR #2288 added WaitForCompletionAsync to DrainAsync() to wait for in-flight messages during graceful shutdown. However, DrainAsync() is also called from within the handler pipeline when a rate-limited message triggers PauseListenerContinuation:

_receiver block execute(message B)
  → pipeline.InvokeAsync()
    → RateLimitContinuation → ReScheduleAsync (stores in Redis sorted set)
    → PauseListenerContinuation → agent.PauseAsync()
      → StopAndDrainAsync()
        → receiver.DrainAsync()
          → _receiver.WaitForCompletionAsync()  ← waits for message B to finish
                                                  but message B is waiting for THIS to return

This circular dependency causes a 30-second deadlock (bounded by DrainTimeout), after which the rate-limited message's retry window has long passed, and the test's 20-second polling timeout expires first.

Fix

Use the _latched flag to distinguish the two call paths:

• Shutdown: OnApplicationStopping calls Latch() first → _latched is already true when DrainAsync() runs → safe to wait
• Pipeline pause: no prior Latch() call → _latched is false when DrainAsync() runs → skip the wait to avoid deadlock

Test plan

• rate_limited_messages_are_delayed_with_native_scheduling passes consistently (was failing every run)
• All 87 Wolverine.Redis.Tests pass
• All 1160 CoreTests pass

Closes #2291

🤖 Generated with Claude Code