perf: optimize MQTT cross-protocol path (0.30x → 0.78x Go)

Replace per-message async fire-and-forget with direct-buffer write loop
mirroring NatsClient pattern: SpinLock-guarded buffer append, double-
buffer swap, single WriteAsync per batch.

- MqttConnection: add _directBuf/_writeBuf + RunMqttWriteLoopAsync
- MqttConnection: add EnqueuePublishNoFlush (zero-alloc PUBLISH format)
- MqttPacketWriter: add WritePublishTo(Span<byte>) + MeasurePublish
- MqttTopicMapper: add NatsToMqttBytes with bounded ConcurrentDictionary
- MqttNatsClientAdapter: synchronous SendMessageNoFlush + SignalFlush
- Skip FlushAsync on plain TCP sockets (TCP auto-flushes)

tests/NATS.Server.Benchmark.Tests/NATS.Server.Benchmark.Tests.csproj

@@ -8,6 +8,7 @@
     <PackageReference Include="coverlet.collector" />
     <PackageReference Include="Microsoft.NET.Test.Sdk" />
     <PackageReference Include="NATS.Client.Core" />
+    <PackageReference Include="MQTTnet" />
     <PackageReference Include="NATS.Client.JetStream" />
     <PackageReference Include="Shouldly" />
     <PackageReference Include="xunit" />