Derive per-query aggregation-heaviness + NDV metadata

### Is your feature request related to a problem or challenge?

Whether dictionary-encoding a column helps depends not just on its cardinality but on how heavily it's aggregated in the query. There's no existing signal that combines both.

### Describe the solution you'd like

Derive metadata that combines NDV with an estimate of aggregation "heaviness" for a given query plan. Heavy aggregation + low NDV = more room for error/benefit if we decide to cast to dictionary arrays early; this signal should be usable by the optimizer.

### Describe alternatives you've considered

_No response_

### Additional context

sub-issue of : #22891
part of epic : https://github.com/apache/datafusion/issues/24111

Start by reading the parent issue #22891 and the epic #24111 to understand the intended optimizer context, then inspect the query-plan metadata areas they point toward. Done means deriving a per-query signal that combines NDV with estimated aggregation heaviness and making that signal usable by the optimizer.

Written by the indexing model from the issue text.