CLIC dumper, PFNetDense updates

[jpata]

• add a dumper for CLIC .slcio files
• minor fix in PFNetDense distance matrix masking
• optional momentum layer separation in PFNetDense with separate_momentum: yes
• CMS-specific input encoding that precomputes sin, cos of angles, log of energy and pt
• additional metrics: per-class recall (confusion matrix diagonal items)
• optionally compute the loss only on signal particles, masking the hits with no associated particle, enabled with sample_weights: signal_only
• make sample_weights configurable for each loss component (e.g. train regression losses masking only the signal particles, while the classification loss using all particles)
• add a --modifier retrain_energy option to modify the configuration to retrain only the energy regression layer (PFNetDense)
• benchmark model training with cms-gnn-dense-focal and the commit dc0dae6
  • performance summary at 300 epochs: indico
  • weights at /eos/user/j/jpata/mlpf/models/cms/2021_07_02/cms-gnn-dense-focal-285ae825.gpu0.local