Re-implement and finish complete eval pipeline for segmentation

[NiWaRe]

For now there is a quick, fast solution for hyper-opt training.
Should either substitute that with the existing (commented out) eval pipeline (+ make it compatible with the current changes and communication with the hyper-opt scripts) or combine it, e.g. verbose -> normal eval with stats_tabel etc., non-verbose -> current version tidied up.

See 8b31aaf for current eval pipeline.