Optimize the num of stats load items for large SQL and wide table

## Enhancement
After the collecting of predicate columns becomes the default, we can change the way to decide which columns' statistics need to be loaded.

The columns that come from the WHERE clause need to read their full stats(NDV, histogram, topn). The columns that come from JOIN keys/group-by clause/order-by/partition-by only need their NDV information.

So we can reduce the size of the stats objects that we need to load to memory.