Improve TopN pruning logic in ver2 ANALYZE

[time-and-fate]

Enhancement

Ref: #32758

In #24906, we introduced an algorithm to prune TopN when building stats to avoid some bad cases. Later in our internal tests, we found that this algorithm may be too aggressive. Sometimes, we expect some samples to be put in TopN but they are pruned here, especially when NDV is low. We need to tune this logic a bit to make it more conservative.