Created by: tensor-tang

According to @Superjomn 's suggestion, the usage should be like this:

  auto main_predictor = CreatePaddlePredictor(config);

  std::vector<std::thread> threads;
  for (size_t i = 0; i < num_threads; ++i) {
    threads.emplace_back([&, i](
      // clone a predictor which is thread-safe
      // and shared the same parameters
      auto predictor = main_predictor->Clone();
      // prepare inputs and output ...
      predictor->Run(inputs, outputs);
    ));
  }
  // join...

So, we should only focus on thread-safe and sharing parameters.