Post recent handler changes, fast count(*) for cluster was broken.

Seeing as we maintain an exact count for ndb, we can easily use this for an optimisation.

With this patch, and use_exact_count DISABLED, we will use the fast way
of getting count(*) but not use the exact count for the optimiser.

With this patch and use_exact_count ENABLED, we will use the fast way of
getting count(*) and use the exact count for the optimiser.