Basic ideas:
1. Store "popular" values in their own buckets.
2. Also store ndv (Number of Distinct Values) in each bucket.

Because of #1, the buckets are now variable-size, so store the size in
each bucket.

Adjust selectivity estimation functions accordingly.