feat(asap-aware-mapping): reuse tighter sketches across accuracy targets

[zzylol]

Summary

Closes #273. Part of #33.

This PR adds AccuracyReconciliationStrategy, a workload-aware replacement strategy that lets otherwise-identical aggregate consumers share the result built for the strictest compatible accuracy target.

For example, two quantile(0.99, x) aggregates requesting epsilon=0.01 and epsilon=0.05 can use one build satisfying epsilon=0.01, instead of maintaining two independent sketches.

Pre-ASAP CSE remains exact: AccuracyTarget still participates in structural equality, and share_common_subtrees is unchanged. Reconciliation is added as an optional Replacement::Rewrite candidate and remains subject to cost-based selection.

Matching and safety

A tighter sibling is eligible only when both aggregates:

• have one accuracy-bearing measure and no HAVING clause;
• use the same reduction, output names, and input subtree;
• have the same AggIntent variant and the same non-accuracy fields (col, q, or k);
• use approximate accuracy targets whose resolved (epsilon, delta) budgets are in a strict Pareto-dominance relationship; and
• expose a provable unique key in the tighter source's output schema.

AccuracyTarget::Exact is excluded because it uses a different implementation family rather than a tighter point on the same approximate-accuracy budget.

The tighter implementation may use any compatible sketch algorithm. CostModel::size_params now explicitly requires returned parameters to satisfy the supplied accuracy budget, including for custom sizing models.

Planning and costing

• Adds ReplacementProvenance::AccuracyReconciliation so these candidates can be identified without inferring semantics from expression shape.
• Prices reconciliation as a read from the tighter sibling's maintained result, rather than as rebuilding the looser target once per consumer.
• Adds the selected looser-to-tighter dependency to the whole-plan reference graph.
• Propagates reconciled consumers into the tighter sibling's effective_consumer_count, after which the tighter MemoGroup propagates uses through its own selected implementation.
• Accuracy dependency edges run strictly from looser to tighter budgets, preserving acyclic topological selection.

Tests

Coverage includes:

• the epsilon=0.01 versus epsilon=0.05 quantile scenario;
• Epsilon/EpsilonDelta Pareto-ordering edge cases;
• exclusion of exact, mismatched-column, and mismatched-grouping cases;
• preservation of exact pre-ASAP CSE behavior;
• unique-key legality for global and without(...) aggregations;
• reconciliation cost behavior for single and shared consumers; and
• propagation of reconciled consumers into the tighter sibling's global selection group.

Verification:

cargo fmt --check
cargo test -p asap-aware-mapping
# 163 passed; 0 failed