MLPF v2 on 12_1_0_pre3

[jpata]

• based on the MLPF model benchmark training at update with larger dataset 1.2.0 [TF] particleflow#84, ACAT2021 benchmark particleflow#92 and New best hyperparameter config particleflow#90
• The final training was carried out by Eric
• switches TF->ONNX (CPU) for inference
• improves the integration by correctly setting the references of the PFCandidates to the originator particles. Puppi and CHS now work.
• data files in https://github.com/jpata/RecoParticleFlow-PFProducer/tree/mlpf_v2_CMSSW_12_1_0_pre3_updated
• mlpf workflows are now 11834.13 (TTbar) and 11843.13 (QCD)
• slides for this PR are here: https://indico.cern.ch/event/1093938/#6-mlpf-in-cms

[jpata]

This was merged in the official CMSSW, so I'm closing it here.