[tmva][sofie] Add overload for inference code that takes output params

[guitargeek]

This makes it possible to avoid memory allocations during inference, also also will help when differentiating through the code with Clad.

Here is a showcase of how the output code of SOFIE would look like with this change. Note that I chose the name doInfer for the new function to better distinguish it from the existing infer method, but I'm happy to rename it to anything including infer as well.

struct Session {
   ...
   // Inference implementation is forward declared:
   void doInfer(float *tensor_x,
                float *tensor_theory_params,
                std::vector<float> &output_tensor_linear_3);
   ...

   std::vector<float> infer(float *tensor_x, float *tensor_theory_params)
   {
      std::vector<float> output_tensor_linear_3;
      doInfer(tensor_x, tensor_theory_params, output_tensor_linear_3);
      return output_tensor_linear_3;
   }
}; // end of Session

void Session::doInfer(float *tensor_x,
                      float *tensor_theory_params,
                      std::vector<float> &output_tensor_linear_3)
{
   // inference code dumped here:
   ...

   // outputs are filled at the end
   using TMVA::Experimental::SOFIE::UTILITY::FillOutput;
   FillOutput(tensor_linear_3, output_tensor_linear_3, 1);
}

[lmoneta]

LGTM!
Thank you Jonas for this contribution!

[github-actions]

Test Results

    18 files      18 suites   4d 3h 51m 37s ⏱️
 2 738 tests  2 736 ✅ 0 💤 2 ❌
47 611 runs  47 609 ✅ 0 💤 2 ❌

For more details on these failures, see this check.

Results for commit 6ccaf8e.