The wide-grid short-window pin flushes pending table stats before it counts reads

[erikdarlingdata]

Refs #4310. Test-only; no product code touched.

Why

QueryStoreIntervalWideGridLiveTests.ShortWindowRead_IssuesNoRoundTripAgainstTheWideTable proves that a
short-window Query Store grid read issues zero round trips against collect.query_store_interval_wide, by
snapshotting pg_stat_user_tables.seq_scan/idx_scan before and after the read under test. It failed once in
CI with Expected (0, 9), Actual (1, 9) and passed on a re-run with no code change.

PostgreSQL throttles a backend's cumulative-stats report to at most once per second
(PGSTAT_MIN_INTERVAL). The test's own setup issues real scans against the wide table from a different
backend (the seed runner's own connection). If that backend's pending report has not flushed yet when the
test takes its "before" snapshot, and flushes moments later, the "after" snapshot picks up a scan the test
under test never issued — a stats-visibility race, not a defect in the code the test exercises.

What changes

Darling/Darling.Tests/QueryStoreIntervalWideGridLiveTests.cs:

• Before the "before" snapshot, force every backend that touched the wide table during setup
  (this test's own connection and the seed runner's pooled data source) to report its pending stats
  immediately with pg_stat_force_next_flush(), so "before" reflects setup's scans completely rather than
  a partially-flushed view of them.
• Before the "after" snapshot, force this test's own connection's stats to report, so a real regression is
  still visible immediately.
• The end-to-end read runs through the viewer's own pooled data source, which this test cannot call
  pg_stat_force_next_flush() on directly. A short delay (1.1s, longer than the once-per-second throttle)
  stands in for a flush on that connection — the documented fallback shape for this class of flake when the
  product read doesn't go through an instrumentable data source.
• No product code changes. The doc comment above the test now explains the flush ordering and why it is
  needed.

Test plan

• Both Darling.Tests and Lite.Tests build 0 errors / 0 warnings.
• ShortWindowRead_IssuesNoRoundTripAgainstTheWideTable run 20 times in a loop against a fresh
  timescale/timescaledb:2.30.1-pg18 container: 20/20 green.
• The whole QueryStoreIntervalWideGridLiveTests class plus DocCommentHygieneTests: 81/81 green.
• A real mutation: temporarily replaced the gate's minimum-window refusal
  (windowEnd - windowStart < minWindow) with false, forcing a short window to route to the table. The
  pin went RED (Expected (1, 9), Actual (1, 10)), proving it still catches the bug it exists for. Reverted;
  clean diff confirmed against the committed source; rebuilt 0 errors.

CHANGELOG entry

SECTION: None
ENTRY: None: test-only de-flake, no user-visible effect.