Implement monitoring for jet quantities

[jpata]

Cluster jets for events in the batch, log plots to comet and tensorboard.

We used to do it in TF, but we didn't implement it really in pytorch, because it requires an extra eval step over some significant sample of events (e.g. 1000 events) with jet clustering, which only runs on CPU currently.

A good approach could be to run the code path under args.test==True (https://github.com/jpata/particleflow/blob/main/mlpf/model/training.py#L1028), essentially run_predictions and make_plots for the hardest sample (e.g. cms_pf_qcd for CMS) for ~500 events, for each epoch.