A tablesample count executes in two phases:
* Estimate table sizes based on reltuples.
* Based on the estimate:
  * If the table is considered 'small', execute an exact relation count.
  * Otherwise, count on a sample of the table using TABLESAMPLE.

The size of the sample is chosen in a way that we always roughly scan
the same amount of rows (see TABLESAMPLE_ROW_TARGET).