load CQ->tail only once during lockless stealing

Currently we load the CQ->tail with acquire semantic to determine
if we should steal from teh victim and load it again in the actual
stealing logic which will also immediately abort if there are no
CQEs to steal.

Keep the optimization for the locked case.

emper/io/IoContext.hpp

@@ -413,6 +413,12 @@ class IoContext : public Logger<LogSubsystem::IO> {
 			return reapCompletionsLockless<callerEnvironment>(continuations, toReap);
 		}
 
+		// CallerEnvironment::EMPER means we are in a stealing worker.
+		// Check if the victim's CQ is empty before taking the lock.
+		if constexpr (callerEnvironment == CallerEnvironment::EMPER) {
+			if (!cqeCount()) return 0;
+		}
+
 		return reapCompletionsLocked<callerEnvironment>(continuations, toReap);
 	}
 

emper/strategies/AbstractWorkStealingScheduler.cpp

@@ -139,14 +139,12 @@ popTop:
 
 	if constexpr (emper::IO_STEALING) {
 		auto* victimIo = runtime.ioContexts[victim];
-		if (victimIo->cqeCount()) {
 		fiber = victimIo->reapSingleCompletion<CallerEnvironment::EMPER>();
 		if (fiber) {
 			emper::statsIncr(awss::stats.nextIoFiberStolen);
 			return std::make_pair(fiber, FiberSource::ioStolen);
 		}
 	}
-	}
 
 	return std::nullopt;
 }