perf(query-engine): avoid cloning native range plan outputs #816

Problem

QueryPlan execution clones every node output for its consumers, and native Estimate clones resolved reads again. For range queries with many groups and buckets, this retains several full bucket-map copies at peak.

Scope

Redesign plan-output ownership so dependency inputs are moved or cheaply shared, for example with Arc, while preserving DAG execution semantics and existing error behavior.

Acceptance criteria

• Native range execution does not clone full read, composed, or resolved bucket maps merely to pass DAG inputs.
• Existing range-query behavior and test coverage remain unchanged.
• Memory ownership is explicit at the plan-runtime interface.