fix(tests): let pytest-asyncio own the suite, drop anyio markers

[ClaraVnk]

Contexte

Test (redis) has been red on main since the toolchain drifted (baseline 29729210049), failing on a fixture teardown rather than a test:

ERROR at teardown of <last test of tests/test_api.py>
RuntimeError: Event loop is closed
  tests/test_api.py:26: await app.state.storage.close()

Every test passed; only the finalizer errored. Test (sqlite) was always green.

Root cause

Two async plugins were driving the same tests. The suite carried @pytest.mark.anyio markers while pyproject.toml sets asyncio_mode = "auto". anyio parametrized the marked tests — hence the [asyncio] suffix in the test IDs — and ran them in loops it managed itself, while the module-scoped client fixture stayed on pytest-asyncio's module loop.

The Redis connection pool therefore accumulated connections opened on per-test loops that were already closed by the time the module fixture tore down. disconnect() then touched dead transports. SQLite never surfaced it because aiosqlite is loop-agnostic.

Note the failing teardown ran on a live loop — the loop the pool's connections belonged to was a different, dead one. That is why "is the loop closed at teardown?" kept giving a misleading answer.

Evidence

The failure never reproduced locally, so I instrumented CI directly (29730652856 before, 29730869139 after):

	fixture setup	teardown	finalization
before	no current event loop	live loop, still errors	no current event loop
after	id=140216260195536	same id	same id

One coherent loop per module, and all three jobs green.

Changements

• Remove the 47 @pytest.mark.anyio markers across 4 test files. asyncio_mode = "auto" already handles every async test, so they were redundant — and they were the thing introducing a second loop owner.
• Document the invariant in pyproject.toml, next to the existing module-scope rationale, so it does not regress.

No production code touched. Test count is unchanged at 114, and the [asyncio] parametrization is gone from the IDs.

Tests

• CI, on the instrumented branch carrying this exact change: Lint, Test (sqlite) and Test (redis) all green, 114 passed.
• Local, CPython 3.14.4 and 3.14.6 (via python:3.14.6-slim), both backends against redis:7-alpine: 114 passed.
• ruff check / ruff format --check clean.

Risques

The real risk is a marker removal silently turning async tests into no-ops. Guarded against: collection still reports exactly 114 tests, and the Redis-backed run genuinely exercises the backend (it is what surfaced this bug in the first place).

Earlier attempt #4 was closed after its diagnosis was falsified; this supersedes it.