Flat-buffer state columns for fixed-size sketches need slice kernels in asap_sketchlib #764

Problem, by example

A 10-second pane of count(distinct ...)-style HLL states, one per series, over 2,000 series at precision 10. #762 makes the executor columnar. Its state column (values::StateColumn) still holds these states as 2,000 separate HllSketchAccumulator objects (StateStorage::Objects), and each object owns its own 1,024-byte register Vec<u8>:

The same applies to Count-Min (rows × cols f64 counters per group) and plain Count Sketch.

Why it isn't done today

The update, merge and estimate logic of these sketches lives in asap_sketchlib, and the executor must use it rather than reimplement it (hashing, register update rules):

• asap_sketchlib::CountMinSketch keeps its counters in a crate-private backend (backend: SketchlibCms). There is no constructor over, or view of, an external counter slice.
• asap_sketchlib::HllSketch exposes registers: Vec<u8>, but its update (hash, leading-zero count, register max) has no function that takes a caller's register slice.

So the executor keeps one sketchlib object per group, which #762 allows as the fallback ("otherwise one object per group").

Proposal

asap_sketchlib exposes, per fixed-size sketch (HLL, Count-Min, Count Sketch), kernels over a caller-owned slice:

• update(slice, item)
• merge(dst_slice, src_slice)
• estimate(slice, ...)
• a slice length for the sketch's parameters

StateColumn then gains a StateStorage::Flat { buffer: Vec<T>, stride } variant. StateColumnUpdater, SummaryMerge and SummaryEstimate call the slice kernels, with no per-group object.

Scope and acceptance

• In scope: HLL, Count-Min and Count Sketch state columns as flat buffers, with the same states (byte for byte) as the per-object path, checked by tests/state_column_updater.rs.
• Out of scope: variable-size sketches (KLL, DDSketch, KMV, UnivMon, the top-k heaps), which stay one object per group.

Status and related work

• asap-executor: a columnar physical operator library (state columns, per-sample updaters) for runtimes executing the DAG #762, phase 3 PR (state columns and the per-sample updater): exact accumulators are fixed-size records there, and sketches are one object per group.
• Needs asap_sketchlib support first; the executor side is small once the slice kernels exist.