fix: Improve consistency of per-column stats on FilterExec output (#22718)

## Which issue does this PR close?

- Closes #22716 

## Rationale for this change

#21081 capped the NDV at the row count when computing statistics for
several operators. This PR extends that work and ensures that per-column
statistics for filter operators are consistent with the estimated output
row count. In particular:

* Null count is also capped at the row count
* Byte size is scaled down by the estimated selectivity

We also extend the analysis to consider null-rejecting predicates; for
example, the clause `a = 10` as a top-level conjunct implies that the
null-count of the surviving rows is exactly 0.

## What changes are included in this PR?

* Ensure per-column statistics (null count, byte size) are consistent
with filtered row count
* Check for null-rejecting predicates to estimate a more accurate null
count of 0
* Update SLT expected plans
* Add unit tests for new behavior
* Various refactoring and comment improvements

## Are these changes tested?

Yes; new tests added.

## Are there any user-facing changes?

No.

Author: neilconway