switch to tfds array_record, improve visualization, dataset descriptions

[jpata]

• update plots for the paper
• switch to ARRAY_RECORD for tfds storage to allow the same datasets to be easily loaded from tf and pytorch
  • this implies new versions of the tfds datasets, propagated to the training configs
• disable tf2onnx as it seems incompatible with the latest tensorflow (due to the flatbuffers requirement)
• reorganize and clean up some older datasets
• switch tests to python 3.10 and ubuntu 22.04

The latest 1.5.0 versions of the tensorflow datasets are available from:

gfal-copy -r root://xrootd.hep.kbfi.ee:1094/local/joosep/mlpf/tensorflow_datasets/clusters
gfal-copy -r root://xrootd.hep.kbfi.ee:1094/local/joosep/mlpf/tensorflow_datasets/hits

and also here at lxplus for the cluster-based dataset:

$ du -csh /eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/*
11G	/eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/clic_edm_qq_pf
11G	/eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/clic_edm_ttbar_pf
13G	/eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/clic_edm_ttbar_pu10_pf
8.6G	/eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/clic_edm_ww_fullhad_pf
4.1G	/eos/user/j/jpata/mlpf/tensorflow_datasets/clusters/clic_edm_zh_tautau_pf

Validation
on the cluster-based dataset, training parameters/clic.yaml, I'm getting