DeepTau v2p5 integration

[ofivite]

Description:

This PR adapts DeepTauId producer to a new v2p5 training. The changes aren't significant since the architecture and data preprocessing hasn't changed much structurally. They include:

• subversion switching between v2p1 and v2p5 and additional process "deepTau2018v2p5" in runTauIdMVA.py
• functionality to drop specific global variables from a general list (TauBlockInputs)
• unification of the scaling interface across subversions and reading of scaling constants from a separate namespace
• corresponding adaptations needed for v2p5, see full list here: Adapt code of DeepTauId producer to be compatible with v2p5 (and still back compatible with v2p1) TauMLTools#99

Backward-compatibility checks:

[v2p1] updated CMSSW code fully reproduces predictions for default (previous) CMSSW
[v2p5] prediction & grid differences for the same model between updated CMSSW and python standalone inference are negligible -- O(1e-5)

P.S. Thanks a lot @kandrosov @shedprog @lguzzi for us working together on this integration!

[mbluj]

@ofivite thank you for the PR. I am starting to revise it. To allow me to run updated code with the 2p5 training, you should also prepare a PR with new training graphs with the trianing to https://github.com/cms-tau-pog/RecoTauTag-TrainingFiles/tree/DeepTau2017v2p5

[mbluj]

@ofivite After a quick review I do not see any significant issues. I plan implement a few small and cosmetics changes myself together with adding the new version of deepTau to the miniAOD workflow. I will keep you informed in case some issues or questions are raised during central review.

[ofivite]

Perfect, thank you @mbluj !