feat(ir): make Aggregate's reduction kind explicit, not inferred (#165)

[zzylol]

Summary

Implements design proposal #165.

QueryExpr::Aggregate previously carried by: GroupKeys alone. Whether a node was a genuine cross-series reduction or a per-series/windowed pass-through with no grouping concept at all (rate, *_over_time, ...) was inferred, downstream, from by.is_empty() && (intent.is_per_series() || child is TimeRange/Subquery) — computed independently in at least two places (output_schema_in, asap-plan::bind) and dependent on the TimeRange/Subquery child staying structurally adjacent, an invariant nothing in the type system protected. A future rewrite pass inserting a node between an Aggregate and its range-window child would silently misclassify a per-series node as a cross-series one — a wrong merge, the same failure mode #163 was about, just relocated one layer up.

This borrows the dbt idea discussed on #165: keep "what kind of computation is this" (a measure's agg type) and "which dimensions to group by" (query-time group_by) as two independent, explicitly-declared things — never let the second answer the first.

Changes

• asap-ir: Reduction::{Reduce(GroupKeys), PerEntity} replaces Aggregate.by. Decided once, at L2→L3 lowering (which knows unambiguously whether it's building a bare per-series range function or a genuine aggregation operator), using the same structural signals the old heuristic used — just computed once, at the producer, instead of re-derived by every consumer.
• output_schema_in/aggregate_output_schema read Reduction directly; the old heuristic and its is_range_child/per_series recomputation are gone.
• asap-plan: bind_summary_agg reads Reduction directly instead of recomputing per-series-ness itself.
• asap-l2: canonicalize's heavy-hitter/ROW_NUMBER rewrites and cse's pass-through rewrite updated to the new shape.
• Every test constructing an Aggregate literal across the workspace updated — most were mechanical, but each Reduce vs PerEntity classification was checked against the actual query shape, not guessed.

Not a code-level bug fix by itself — today's construction sites already got this right — it closes the door on the next one, by making the type system enforce what used to be an unenforced convention.

This lands independently of #164 (the L4-level SummaryGrouping fix for #163); once merged, #164 can be simplified to read this field directly instead of re-deriving the same judgment a second time via its own heuristic.

Test plan

• Every Reduce vs PerEntity classification verified against the existing PromQL/SQL conformance suites — 100/100 in promql_conformance.rs, 44/44 in promql_lowering.rs, full SQL corpora — rather than asserted from first principles alone.
• cargo build --workspace / cargo test --workspace — clean, zero failures across every crate.
• cargo clippy --workspace --all-targets --all-features -- -D warnings — clean.
• cargo fmt --all --check — clean.

🤖 Generated with Claude Code