test(group): oracles for the fixes in #540 and the coverage gaps

[ser-vasilich]

What & why

Tests for #540, targeted at its branch because they pin behaviour that only exists there.

The four fixes this branch carries, each red on the merge-base cb6b0371 and green here (f92150bc):

• group/topk_full_size_k.rfl — top/bot with K ≥ the group's value count came back in heap order (bot v 1024 over [5 3 9 1 7] → [1 3 7 9 5]); sorted output is asserted, with and without nulls thinning the group, and on a 200k-row group for the parallel consumer. The null case is the "serial heap not heapified when nulls leave fewer than K values" fix seen from the query language.
• group/count_distinct_str_many_groups.rfl — count (distinct text) over 65536 groups with a four-word vocabulary crashed with SIGSEGV; every group's distinct count is checked against its row count.
• group/list_key_count_distinct.rfl — LIST-typed keys with count (distinct v) beside a streaming aggregate failed with vec_new: … got LIST.

Coverage the review found missing (7f49fd2d):

• group/first_last_parallel_oracle.rfl — first/last on a 280k-row group that the two-worker suite splits at the parallel grain, checked against the rows themselves rather than another grouping path, with a nullable column and a selection.
• group/topk_k_limits.rfl — K admitted for 1 and 1024, rejected for 0, −1 and 1025, on a small table and past the parallel threshold.
• test_wide_count_distinct gains a table whose distinct count depends on the key's parity, with expectations read from the emitted key column, so a count landing on the wrong key is detected; the uniform table asserts the same value for every group and cannot tell.

All six pass in the ASan runner on 19154617.

Checklist

• PR targets dev (not master) — targets perf/grouping-engine-scaling on purpose; four of the tests fail or crash on dev
• Commits follow Conventional Commits (feat: / fix: / perf: / docs: / …)
• make builds cleanly (no new warnings)
• make test passes; tests added/updated for behaviour changes