video_core/gpu_thread: Keep the write lock for allocating the fence.

Else the fence might get submited out-of-order into the queue, which makes testing them pointless.
Overhead should be tiny as the mutex is just moved from the queue to the writing code.

src/video_core/gpu_thread.cpp

@@ -151,11 +151,13 @@ void ThreadManager::OnCommandListEnd() {
 }
 
 u64 ThreadManager::PushCommand(CommandData&& command_data) {
+    std::unique_lock lk(state.write_lock);
     const u64 fence{++state.last_fence};
     state.queue.Push(CommandDataContainer(std::move(command_data), fence));
 
     if (!is_async) {
         // In synchronous GPU mode, block the caller until the command has executed
+        lk.unlock();
         WaitIdle();
     }
 

src/video_core/gpu_thread.h

@@ -101,7 +101,8 @@ struct CommandDataContainer {
 struct SynchState final {
     std::atomic_bool is_running{true};
 
-    using CommandQueue = Common::MPSCQueue<CommandDataContainer>;
+    using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
+    std::mutex write_lock;
     CommandQueue queue;
     u64 last_fence{};
     std::atomic<u64> signaled_fence{};