perf(query): narrow + materialize before group for WHERE + multi-key

[ser-vasilich]

Summary

The pre-filter narrowing path in ray_select_fn previously fired only
when the by-dict contained a computed val (e.g. q42's
(xbar EventTime ...)). For bare-ref multi-key shapes with a
selective WHERE — e.g. ClickBench q30 / q31 — the planner left the
work to the fused mk_par_v2 path. That path reads each by-key + agg
input column from the original wide table at the sparse positions
left by the WHERE bitmap. On the 100-column hits table with ~14%
selectivity, the gather wastes a cache line per touched column per
passing row.

Narrowing the input down to just the referenced columns and filtering
once gives the downstream group a dense column-store and skips the
gather.

Extend the gate to fire when:

• the desc:/asc: COUNT take N shape matches,
• by-vals are all bare column refs (no computed val),
• the by-dict has ≥ 2 keys, and
• at least one aggregate has an input column distinct from the
  by-keys (sum / min / max / avg, not pure count).

The count-only-on-a-by-key shapes (q14: count SearchPhrase by
{SearchEngineID, SearchPhrase}; q40: count URLHash by
{URLHash, EventDate}) reuse the by-key column for the count input
— narrowing costs the projection without saving the gather, so the
gate keeps them on the original fused path.

ClickBench 10M (on top of the gate from #227):

q30  ~152 →  ~50 ms   (-67%)
q31  ~353 →  ~55 ms   (-84%, -298ms)

Full 43-query sum (with #227 already in): -394 ms / -11.3%.

Tests: 3232/3234 pass (unchanged).