Timing: a node shared across batch queries with different timings fails with ConflictingTiming #654

Found in the #541 review.

Problem

apply_lifecycle_timings with one TimingMemo across a batch fails with ConflictingTiming when two queries share a node but need different timings. Example:

• query 1 maintains SummaryAgg(exact sum, scan) at ingestion time;
• query 2 reads the same Scan at query time through Limit(scan).

The cause: split_shared_by_phase splits shared nodes only within one root, while the memo spans every root of the batch. The batch test (types/tests/physical_export.rs) only shares a node at a single timing, so it does not cover this case.

Why it is latent

At the top of the stack, the only workload-timing caller (execution_timed_dag) uses MaterializationAssignment::all_query_time(), so a shared node is always timed the same way.

Follow-up

When ingestion-time maintenance comes back, add a workload version of split_shared_by_phase that takes &[Rc<OperatorNode>]. It should collect the timings reached from all roots and rebuild them with one shared copies map. Add a two-query regression test.