MVA generated code is not thread-safe (Follow-up from "PrLongLivedTracking: remove use of `static` in MLP")

The following discussion from !1053 (merged) should be addressed:

• @graven started a discussion:

  Aside: PrLongLivedTracking is not reproducible 'as is' when run multithreaded. The number of tracks produced varies from run to run. As the code it self looks thread-safe (after I changed the counters to the new Gaudi counters), I tried the following:

  @@ -485,7 +492,12 @@ LHCb::Tracks PrLongLivedTracking::operator()(const LHCb::Tracks& InputTracks) co
           m_debugTool->tuneFinalMVA( track.track(), trueTrack, vals);
         }
   
  -      const double mvaVal = m_mvaReader->GetMvaValue( vals );
  +      auto mva = [&](auto& v) {
  +         std::lock_guard<std::mutex> guard(mvaMx);
  +         return m_mvaReader->GetMvaValue( v );
  +      };
  +
  +      const double mvaVal = mva( vals );
         track.setMVAVal( mvaVal );
   
         foundBestCand = foundBestCand || mvaVal > 0.05;
  

  where mvaMx is a mutex defined in the global scope. And lo-and-behold, the result becomes reproducible. So the TMVA generated code is not thread-safe in more ways... (how can just asking the evaluate an (already initialized) MVA be not thread-safe???)