feat(sketch_algebra): unify SketchKind onto asap_sketch::SummaryKind

[zzylol]

Summary

Step 3 of the sketch-identity unification (design doc, Steps 1–2: ASAPController#141/#142, PR #399). Replaces sketch_algebra::{SketchKind, SketchParams} (5 approximate-sketch-only variants, heap-bearing as a with_heap params flag) with asap_sketch::{SummaryKind, SummaryParams} (14 variants: exact accumulators + approximate sketches, heap-bearing promoted to kind identity) across all 24 real consumers, then deletes the now-empty sketch_params.rs entirely.

Upstream: bumped ASAPController's pin to 150ef7d (merge of PR #142, "derive PartialOrd/Ord for SummaryKind") — needed for BTreeSet<SummaryKind> in the 5-sketch routing-connector emit path's deterministic-order contract. No serde needed upstream — confirmed by audit that every real emit path builds JSON/YAML by hand, never via serde_json::to_value on a whole struct. The existing Serialize/Deserialize derives on PhysicalExpr/BackendAggregation/EdgeSketchProcessor/GatewayMergeProcessor only ever backed #[cfg(test)] round-trip tests, now deleted (3) or rewritten to check Debug-string substrings / clone-equality instead (2), preserving what they actually verified.

Behavior-preservation notes (not a pure mechanical rename)

• Wire text is unchanged. build_backend_aggregation_json's JSON shape is byte-identical — sketch_kind_to_backend_type/sketch_params_to_json read the new shape but produce the same strings/fields, including the pre-existing quirk where Cms's JSON never carried a with_heap key (only CountSketch's did) and heap_size was never emitted on this wire path at all.
• Heap-bearing dispatch. Every function that used to match a bare SketchKind (Cms/CountSketch) regardless of with_heap now matches Cms | CmsWithHeap / CountSketch | CountSketchWithHeap so topk-bound aggregations keep routing through the same processor names, merge processors, and warm-shape classification as before.
• The 5-sketch routing-connector path (FAMILY_ORDER, needed_families, family_to_proc) is bare-5-family-keyed by design — added a base_family() normalizer so a committed heap-bearing kind still matches its bare FAMILY_ORDER entry instead of being silently dropped. This fixes a real latent bug: without normalization, a metric committed to a heap-bearing family would have silently vanished from this emit path.
• bind_cms_topk.rs now threads the real requested top-k k into heap_size (previously unrepresentable). The backend wire path doesn't read this field either way, so this is strictly more correct in-memory data with no observable behavior change.
• SketchType <-> SummaryKind conversions moved to types.rs, next to SketchType's own definition (orphan rule).
• Test expectations updated (not regressions) — several tests asserted a bare CountSketch/Cms kind for a topk-bound aggregation; the underlying behavior (heap-bearing wire params) was already correct before this change, only the enum value identifying it changes.

Test plan

• cargo build --workspace — clean
• cargo test -p control_plane — 822 passed, 1 pre-existing unrelated failure (invalid_sketch_type_override_falls_back_to_default)
• cargo test -p data_plane --lib — 881 passed, unaffected

🤖 Generated with Claude Code