CSE detection should run in pre-ASAP IR; post-ASAP can run a secondary pass

[zzylol]

Summary

Resolves the "Where it runs" open question in #212. CSE detection (structural
recognition of shared sub-computations within a query, or across a
QueryWorkload batch) should be implemented as a pass over the pre-ASAP
QueryExpr IR (asap-types::pre_asap), not inside asap-aware-mapping.
Post-ASAP binding can optionally run a second, narrower CSE pass over the
already-implemented SummaryExpr/SummaryNode DAG, but that is a
secondary optimization on top of the primary one, not where detection should
mainly happen.

Why pre-ASAP, not post-ASAP

1. Batch query optimization needs to see shared work before summary
   binding, not after. This is motivated by Peilin's batch query
   optimization work: batching only pays off when the planner recognizes
   duplicate/overlapping computation before deciding how each piece gets
   realized. Recognizing it only after realization means summary choices for
   each query are already made independently, and reconciling them after the
   fact is a strictly harder problem (summary-level equivalence/subsumption
   reasoning) than sharing the pre-ASAP subtree that feeds a single bound
   summary in the first place.
2. The legality primitive already lives at this layer. Schema::unique_keys
   / is_reusable() is a pre-ASAP schema property, and its own doc comment
   already says it "feeds CSE's producer-sharing legality check" — the IR
   was already built expecting detection to happen here.
3. asap-aware-mapping::implement_tree runs per query today. If CSE only
   existed post-implement, every query would still get bound to its own
   summary independently, and CSE would have to retroactively merge two
   already-built SummaryAgg nodes instead of sharing one pre-ASAP subtree
   before either is built.
4. Consistency with "normalize semantics, not syntax" (docs/pre-asap-ir.md)
   — CSE is a normalization-adjacent concern (recognizing equivalent
   computation) and belongs next to canonicalize, which already runs at
   this layer.

Where post-ASAP CSE still has a role

Not every sharing opportunity is visible at the pre-ASAP level. Two
structurally different pre-ASAP nodes — e.g. Quantile(x, 0.99) and
Quantile(x, 0.95) — can legitimately share one built sketch, read out
twice. That's only recognizable once both are bound to the same
(SummaryKind, params) at post-ASAP, since pre-ASAP structural equality
(which includes the accuracy target on AggIntent) will correctly treat them
as different nodes. So the design is two passes, not one:

• Primary — pre-ASAP structural CSE. Over QueryExpr, before
  implement, keyed on structural/value equality, gated by unique_keys.
• Secondary — post-ASAP summary-level CSE. Over already-implemented
  SummaryNodes, keyed on (SummaryKind, params, source-subtree) equality,
  recognizing readout-level sharing the pre-ASAP pass can't see by
  construction.

Scope of this issue

This issue settles the placement decision and updates #212 accordingly. The
detection algorithm itself is tracked in a follow-up design-doc issue. The
prerequisite representation work — pre_asap::QueryExpr needs a
shared-subtree representation (Rc instead of Box) before any CSE pass can
express sharing at all — is being landed as mechanical prep ahead of that
design doc.

Related

• New feature: add common sub expr elimination #212 — parent issue; this resolves its "Where it runs" open question
• crates/asap-aware-mapping/src/lib.rs module doc (names "which shared
  sub-expressions to hoist" as a stated but unimplemented job)
• crates/types/src/pre_asap/schema.rs (unique_keys / is_reusable)

[zzylol]

@milindsrivastava1997 @Selvomega