GModel based post-processing datasets use standard vectors for indexing
This should be documented, as this can lead to unexpected behaviour (see e.g. #346 (closed)) when the entity numbering is not dense.
Applying the same trick as for mesh caches (using a map) leads to poor performance.