test(compat): autouse fixture + cross-adapter dedup test for brand_manifest warning

[bokelley]

Summary

Fixes #689. Test-hygiene follow-ups flagged on the #688 review.

• Autouse fixture in tests/conftest.py clears adcp.compat.legacy.v2_5._url._brand_manifest_path_warned between every test. The 10 existing warning tests drop their monkeypatch.setattr(...) boilerplate. A future test that forgets the reset can no longer flake on inherited state.
• New test_brand_manifest_dedup_is_cross_adapter documents the shared-state semantic introduced in Extend brand_manifest non-standard-path warning to inline-object branch #687 — same offending URL through get_products then create_media_buy warns exactly once per process. Catches a regression back to per-module dedup.

Verification

• pytest tests/test_legacy_adapter_v2_5_*.py — 65 passed (was 64; +1 cross-adapter test, no net loss from the removed monkeypatch calls).
• Manually simulated the per-module regression (_brand_manifest_path_warned.clear() between adapter calls): two warnings emitted, confirming the new test would catch that change.
• ruff check + mypy clean.

Out of scope

Other nits flagged across this cluster but not addressed (intentional — small or theoretical):

• Bounded dedup cache (_brand_manifest_path_warned is an unbounded set). Per-deployment cardinality of distinct brand_manifest URLs is small in practice; revisit if a real growth signal appears.
• strip_url_scheme rename (now a bare-domain fallback for extract_brand_domain) — cosmetic.
• IPv6 silent failure on extract_brand_domain ([::1] → "::1" fails BrandReference.domain regex without typed error) — theoretical for brand_manifest URLs in practice.

[bokelley]

Expert review

code-reviewer: approved — no blockers. Confirmed:

• .clear() vs monkeypatch.setattr(..., set()) semantic change is safe (state is only referenced via attribute lookup inside _url.py; no code caches the set by reference)
• Cross-adapter test is correctly placed with an explicit regression-mode assertion message
• Direct attribute access for the test-only reset is preferable to expanding the public API with a clear_dedup() helper

Nit (non-blocking): the autouse fixture lives at top-level conftest.py so it runs .clear() on every test in the suite. Cost is negligible; a scoped sub-conftest would be more surgical if compat tests ever get their own subdirectory.

CI: 15/15 green. Merging.