[Chore] Replace the multi-group Measure row merger with a batch merger

Summary

Merge vectorized Measure results without first converting them into individual InternalDataPoint objects. Follow-up to #14066 under #13998.

Size audit and dependencies

Classification: tracking parent, not an executable leaf or automated implementation task. Boundary: the production multi-group Measure caller of MergeGroupMIterators and its sortableDataPoints / sortedMIterator stack. This spans batch consumption, caller/egress migration, schema alignment, ordering/version deduplication, and resource ownership. Decompose after the legacy removal merges. Blocked by: #14066 via BanyanDB PR apache/skywalking-banyandb#1326.

Requirements

R1. Consume and merge per-group batches directly; do not reconstruct the row merger internally. R2. Preserve requested ordering, version-based deduplication, schema alignment, and pagination. R3. Materialize protobuf only where the response boundary requires it; preserve tracing and response compatibility. R4. Release consumed batches correctly on completion, cancellation, and errors.

Concrete oracle

Two groups supply these records, with duplicate versions split across batches:

timestamp=10, SID=7, version=1, value=100
timestamp=10, SID=7, version=3, value=300
timestamp=20, SID=8, version=1, value=200

Ascending output is (10,7,3,300) followed by (20,8,1,200). Descending reverses that order. Ascending with offset 1 and limit 1 returns only (20,8,1,200).

Completion criteria and RED-test direction

• The real processor multi-group caller consumes the batch-native merger.
• Tests assert the exact oracle across batch boundaries, plus index ordering and supported schema alignment.
• A path-sensitive production-caller regression fails if row iteration/materialization is restored. Output assertions alone are insufficient: the existing merger already produces these answers.
• Each executable child identifies its public boundary and failing test before implementation; no helper-only foundation PR counts as completion.
• Cancellation/error tests verify resource release; benchmark comparisons report allocations without claiming a new resource bound that has not been proved.

Packages: banyand/query, pkg/query/logical/measure, pkg/query/vectorized/measure.

go test ./banyand/query ./pkg/query/logical/measure ./pkg/query/vectorized/measure/...
go test ./test/integration/standalone/query/... ./test/integration/distributed/query/...

Out of scope

Changing ordinary Top or GroupBy semantics; the internal TopN pre-aggregation RPC; schema-evolved aggregate partial reduction (separate workstream).

Parent: #13998.

Compatibility and review policy

• Preserve existing on-disk and wire formats and mixed-version compatibility. Any necessary protocol change requires a separately reviewed compatibility contract.
• Use explicit, hand-calculated expected results; do not depend on the removed row engine as the test oracle.
• Human review enforces the no-new-row-path policy. A row-call-site lint guard is not required.