Make CostCalculator to calculate cumulatively

[sopel39]

This enables:

• incorporating peak memory computation as part of CostCalculator
  (and not CachingCostProvider which has different responsibility)
• simplifying CostProvider and StatsProvider into single
  interface, e.g: TraitProvider<T>
• simplifying CachingCostProvider and CachingStatsProvider into
  single class, e.g: CachingTraitProvider<T>

This should enable cleaner approach for #11591

[findepi]

@sopel39 i encourage you to rebase your refactor on top of #11591. I have already spend quite some time in #11591 and don't want to rework all this.

[sopel39]

TraitCalculator could also be extracted:

interface TraitCalculator<T> {
    T calculateTrait(
            PlanNode node,
            TraitProvider<T> sourceTraits,
            Lookup lookup,
            Session session,
            TypeProvider types);
}

TraitProvider<T> TraitProvider#of(TraitCalculator<T> calculator, ...)

// with caching
TraitCalculator<T> calculator = SomeCachingTraitCalculator.of(calculator);
TraitProvider<T> provider = TraitProvider.of(calculator);

Note: TraitCalculator could use trait rules to compute desired traits.

Memo should be caching decorator of TraitCalculator, so that it can purge entries when groups are changed:

interface Memo {
  ...
  TraitCalculator<T> cache(TraitCalculator<T> calculator, boolean groupTrait);
  ...
}

When groupTrait==false then cached traits for parent groups are purged when group is changed.