Compute metrics during evaluation, not during training

## Context
Currently we are computing all metrics during the model training. This is true both for:
- Training of TMD and DL classifiers:
https://github.com/BlueBrain/morphoclass/blob/35bbe615bbf3a0f085e2d2ca1ebcf67c4d95a868/src/morphoclass/console/cmd_train.py#L139-L141
- Training of Morphometrics classifiers:
https://github.com/BlueBrain/morphoclass/blob/35bbe615bbf3a0f085e2d2ca1ebcf67c4d95a868/src/morphoclass/console/cmd_morphometrics.py#L198-L200

## Actions
- [ ] Make sure that nowhere else in the code (e.g. during evaluation) we are assuming that the `checkpoint.chk` produced during `training` contains metrics scores.
- [ ] Move the computation of metrics scores from `training` stage to `evaluation` stages.

## See also
https://github.com/BlueBrain/morphoclass/pull/45#issuecomment-1115874293