Currently, the child of a SargableNode is not always a ScanNode after the substitution phase. This happens when the SargableMerge fails to merge two SargableNode (e.g. one of them has non-singleton disjunction requirement) in the substitution phase. Such SargableNode will remain in memo until the exploration phase.

Due to this, a SamplingEstimator may encounter such SargableNode and has to fall back to heuristic estimation without an additional check.
 
To address that:

1. Create a SubstituteConvert<SargableNode> rewrite which converts SargableNode back to a FilterNode like what ExploreConvert<Sargablenode> does today by calling lowerSargableNode(sargableNode, ctx).
2. Update the TODO in the sampling_estimator.cpp into an assertion such that the child of a SargableNode is always a ScanNode.