Converge grouped *_over_time onto positional Aggregate.by (residual of review D)

[zzylol]

Context

PR #5 (review D) converged the dominant PromQL grouped-aggregate patterns onto a positional Aggregate.by — the same shape SQL GROUP BY produces:

• sum by(job)(m) (instant vector) → Aggregate { by: [job], [Sum] }
• sum by(job)(rate(m[5m])) → Aggregate { by: [job], [Sum], child: Aggregate{[Rate]} } (the inner rate is modeled as a per-series label-preserving reduction)

This removed the name-based Partition-as-grouping shape (and the "documented droppable L5 hint, yet load-bearing" hazard) for those cases.

What is left

Grouped window reductions — sum by(x)(avg_over_time(m[5m])), max by(x)(quantile_over_time(...)), topk by(x)(avg_over_time(...)), etc. — still fall back to the legacy name-based Partition wrapper. Two coupled reasons:

1. Shared AggIntent variants. avg_over_time (per-series) and cross-series avg both lower to AggIntent::Avg; sum_over_time and sum both to AggIntent::Sum. So per-series-ness can't be read off the intent variant — only off structure (the enclosing time Window).
2. Window hoisting. The converter's canonical-shape logic hoists the *_over_time Window above the outer cross-series aggregate, separating the Window from its Aggregate — so a "Window-over-Aggregate ⇒ label-preserving" rule in schema derivation doesn't see them adjacent. Folding keys into the per-series by naively also breaks downstream sample-value resolution (e.g. topk by(h)(avg_over_time(…)) — fixed in PR feat(promql+sql): unified positional L3 intent algebra (PromQL & SQL lowering) #5 by gating convergence to non-windowed reductions).

Acceptance

• sum by(x)(avg_over_time(m[5m])) → positional Aggregate { by: [x], [Avg], … } (no Partition), with x resolved against a label-preserving per-series window reduction.
• The same for *_over_time under topk/bottomk/count.
• Once all grouped PromQL aggregates are positional, Partition can be reserved strictly for the L5 sharding hint (re-document) and PartitionKeys retired from the grouping path.
• Full PromQL + SQL suite green; clippy -D warnings + fmt clean.

Likely approach: distinguish per-series window reductions structurally (don't hoist the window above the outer aggregate, or carry an explicit per-series marker) so the inner reduction preserves labels — then the existing converter convergence path applies unchanged.

Follow-up to #5 (feat/promql-l1-l3).