we should consider Fields.getUniqueTermCount,
which has a default implementation that sums across fields (preflex overrides).
This was necessary to have some way to access the segment-level unique term count for 3.x indexes,
which do not actually know this information per-field and override this method to provide it.
But there is no need to have this on AtomicReader (I think its way too expert, just get the Fields
and get it from there), and we can consider deprecating this in 4.x and removing it in trunk
alltogether as then someone can just use the field-level statistics.