fix(sort,query): radix key width across dispatch tasks; keep aggregate columns on zero groups

[ser-vasilich]

What & why

Two long-standing wrong-result bugs, each with a test that fails on dev and passes here.

asc/desc/xasc return an unsorted vector. compute_key_nbytes derives the radix key width by letting every dispatch task OR together keys[i] ^ keys[start] — the differences against the task's own first element — and covers cross-task differences with only nw-1 sampled rows, assuming workers own contiguous chunks. The pool dispatches 8192-row tasks, so a high byte that changes only between tasks is dropped from the key width and the sort silently orders by the low bytes alone. 200000 values where rows 40960..49151 carry bit 16 come back as [0 0 0 0 0 65536 0 …]; at -c 1 any nearly sorted vector past 65536 rows with one swap sorts by its low 16 bits. Present since the v2 import (4a3a0c7). Every task now diffs against keys[0], the reference the serial path already uses, which covers every bit exactly; the sampling loop goes. Sort throughput is unchanged on random input (10M rows on the pool 53.6 → 52.7 ms), and an already sorted 10M-row vector at -c 1 drops from 201 to 50 ms because the key width is now right.

A grouped select that matches no row loses its aggregate columns. The eval-fallback grouping paths create each result column on the first group: streaming aggregates allocate agg_vec from the first value's type, and nonagg_eval_per_group_core only creates its result inside the per-group loop. With zero groups both stay NULL, so a single-key select silently dropped those columns (17 requested, 2 returned) and a composite key failed with per-group projection evaluation failed. The expression is now probed once over an empty slice (projections) or the aggregate over an empty vector of the source type (streaming, at all three sites) and an empty column of that type is emitted; when the probe cannot run an empty LIST column stands in. aggr_unary_per_group_buf's I64 guess takes the same probe.

Tests: test/rfl/sort/asc_key_width.rfl, test/rfl/group/zero_groups_keep_columns.rfl. On this branch make test (ASan/UBSan) 3825/3825, make tsan-test 20/20.

Checklist

• PR targets dev (not master)
• Commits follow Conventional Commits (feat: / fix: / perf: / docs: / …)
• make builds cleanly (no new warnings)
• make test passes; tests added/updated for behaviour changes