Cluster error parameterization

Current approach:

• O2/GPU/GPUTracking/Merger/GPUTPCGMTrackParam.cxx:
  
  prop.GetErr2(err2Y, err2Z, param, zz, cluster.row, clusterState, cluster.sector, time, invAvgCharge, invCharge);
  #ifndef GPUCA_GPUCODE
  fprintf(fpdumperr, "%d,%d,%f,%f,%d,%f,%f,%f,%f,%f,%f,%f,%f,%f,%f,%f,%f,%f\n", iTrk, cluster.num, err2Y, err2Z, clusterState, xx, yy, zz, mP[0], mP[1], mP[2], mP[3], mP[4], mC[0], mC[2], mC[5], mC[9], mC[14]);
  #endif
• For every dumped cluster: Parse the cluster.num coloumn and only keep the latest cluster
• Training the NN:
  • labels_x = ["clusterState", "xx", "yy", "zz", "mP[2]", "mP[3]", "mP[4]", "mC[0]", "mC[2]", "mC[5]", "mC[9]", "mC[14]"]
    labels_y = ["yy", "zz", "mP[0]", "mP[1]"]
  • data_Y[0] = data_Y["yy"]**2 - data_Y["mP[0]"]**2
  • data_Y[1] = data_Y["zz"]**2 - data_Y["mP[1]"]**2
• Normalized mC[0] and mC[2] using log10() because their values where O(10^4) to O(10^7)
• Training pipeline is set up but training doesn't converge well yet...

Checking the distributions