fix(memory): restart the cleanup task on a new event loop and add aclose()

[allen0099]

Closes #181

Problem

MemoryBackend started its cleanup task on the event loop of the first cache call and never looked at it again. If that loop was closed without cancelling the task (a manually managed loop, some test runners, a server restart inside one process), the task stayed "pending" forever. Every later call saw a task that was not done, so no cleanup ran on the new loop and expired entries piled up.

Change

• _ensure_cleanup_started() checks that the existing task belongs to the running loop. If not, it cancels the old task and starts a new one on the current loop.
• Cancelling goes through a small _cancel() helper: a task on a closed loop is left alone (cancelling it would raise), a task on the running loop is cancelled directly, and a task on another open loop is cancelled with call_soon_threadsafe.
• New async def aclose(): cancels the task and waits until it has finished, for use in a lifespan shutdown. It only waits for a task on the running loop.
• stop_cleanup() stays synchronous and unchanged in behaviour. It now uses the same helper, and its docstring points to aclose().
• Docs (en and zh-TW BACKENDS.md): a paragraph on the loop restart and a lifespan example with aclose().
• CHANGELOG: an Added entry for aclose() and a Fixed entry for the restart.

Tests

Four new tests in tests/backends/test_memory.py:

• the task restarts on a new loop after the old loop was closed;
• moving loops cancels the task on a loop that is still open;
• aclose() waits for the task;
• aclose() only cancels, and does not wait for, a task on another loop.

Each fix point was mutation-checked by reverting it: ignoring the loop, removing the closed-loop guard, not waiting in aclose(), and not cancelling across loops. Each revert fails the matching new tests.

Local checks: ruff, mypy --strict, the full suite against live Redis and Memcached (882 passed), and zensical build --strict for both languages.