Simplify color algebra (A-iB)(M)(A+iB) as AMA+BMB

[valassi]

This is a spinoff of #155 about color algebra optimisations.

While discussing with @roiser we realised that one can rewrite (A-iB)(M)(A+iB) as AMA+BMB in the color algebra calculation (where A,B are vectors and M is a matrix, with ncol and ncol*ncol dimensions).

This will in any case simplify tensor core calculations (as there is no need to deal with complex arithmetics, one can just use two real quadratic forms).

I was also naively expecting that this might improve performance by a factor 2 or 4, because some calculations are not needed. It turns out that the physics results are the same, but performance is excatly the same.

I would imagine that this may be because in the older implementation the c++ compiler was in any case optimising away some calculations, refusing to compute calculations whose results are not stored? Essentially the change is this

madgraph4gpu/epochX/cudacpp/gg_ttgg/SubProcesses/P1_Sigma_sm_gg_ttxgg/CPPProcess.cc

Line 1942 in 6f5688b

// Sum and square the color flows to get the matrix element

      // Sum and square the color flows to get the matrix element
      // (compute |M|^2 by squaring |M|, taking into account colours)
      fptype_sv deltaMEs = { 0 }; // all zeros https://en.cppreference.com/w/c/language/array_initialization#Notes
      for( int icol = 0; icol < ncolor; icol++ )
      {
        cxtype_sv ztemp_sv = cxzero_sv();
        for( int jcol = 0; jcol < ncolor; jcol++ )
          ztemp_sv += cf[icol][jcol] * jamp_sv[jcol];
        // OLD implementation (slower?)
        //deltaMEs += cxreal( ztemp_sv * cxconj( jamp_sv[icol] ) ) / denom[icol];
        // NEW implementation (faster?) 
        // Rewrite the quadratic form (A-iB)(M)(A+iB) as AMA - iBMA + iBMA + BMB = AMA + BMB!
        deltaMEs += ( cxreal( ztemp_sv ) * cxreal( jamp_sv[icol] ) +
                      cximag( ztemp_sv ) * cximag( jamp_sv[icol] ) ) / denom[icol];
      }

This is almost a no-op but I would merge it anyway. Maybe I will first rerun a full set of tests just for good habit.

I only tested one log and essentially there is no difference with upstream/master:
ed0890b

[valassi]

All tests pased, I am self merging