Adapt code of DeepTauId producer to be compatible with v2p5 (and still back compatible with v2p1)

[kandrosov]

No description provided.

[ofivite]

Subtasks:

• Switch to Run3 test sample and corresponding global tag
• Add subversioning functionality
• Test compatibility with TF graph for v2p5 training
• Check data preprocessing compatibility between v2p1 and v2p5
  • tau_n_charged_prongs: make sure that it is shifted by 1 (mean=1) in the scaling json for v2p5 to account for difference in definitions without changing the interface in DeepTauId.cc: v2p1 -> decayMode / 5 + 1 , v2p5 -> decayMode / 5
  • tau_footprintCorrection: v2p1 uses tau->tauID("footprintCorrectiondR03"), while v2p5 uses tau->tauID("footprintCorrection")
  • tau_flightLength_sig: v2p1 -> 0.55756444, v2p5 -> tau->flightLengthSig()
  • Add a check ndof>0 when filling PF chi2 and ndof, both for v2p1 and v2p5. NB: this will break backward compatibility with python v2p5
  • track-related vars for PF cands: v2p1 -> pseudoTrack(), v2p5 -> bestTrack() -> have checked that there is no difference between best vs pseudo, will leave pseudoTrack()
  • Add dropping functionality and remove tau_phi, tau_dxy_pca_x, tau_dxy_pca_y, tau_dxy_pca_z for v2p5
  • Swap order of PF_e and Electron blocks for v2p5 w.r.t. v2p1
  • Implement NaN preprocessing as in DataLoader: if isfinite -> scale, clamp, fill; else -> fill with 0. NB: this might break backward compatibility with CMSSW v2p1
• Scaling
  • Implement reading from vectors from separate header
  • Unify interface for scaling application between v2p1 and v2p5
  • Rederive scaling params for v2p1 and switch to unified ScalingParams::scale() interface
    - TauFlat
    - GridGlobal
    - PfCand_electron
    - PfCand_muon
    - PfCand_chHad
    - PfCand_nHad
    - PfCand_gamma
    - Electron
    - Muon
• Set disable_hcalFraction_workaround and disable_CellIndex_workaround to true in TauIDEmbedder