Set once_flag to thread_local for StreamSafeCUDAAllcation (#49104)

paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc

@@ -266,7 +266,7 @@ uint64_t StreamSafeCUDAAllocator::ProcessUnfreedAllocationsAndRelease() {
   return underlying_allocator_->Release(place_);
 }
 
-std::once_flag StreamSafeCUDAAllocation::once_flag_;
+thread_local std::once_flag StreamSafeCUDAAllocation::once_flag_;
 
 std::map<platform::Place, std::vector<StreamSafeCUDAAllocator*>>
     StreamSafeCUDAAllocator::allocator_map_;

paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h

@@ -45,7 +45,7 @@ class StreamSafeCUDAAllocation : public Allocation {
   gpuStream_t GetOwningStream() const;
 
  private:
-  static std::once_flag once_flag_;
+  thread_local static std::once_flag once_flag_;
   void RecordGraphCapturingStreams();
   void RecordStreamWithNoGraphCapturing(gpuStream_t stream);
   DecoratedAllocationPtr underlying_allocation_;