fix(memory): stop listing and counting expired entries

[allen0099]

Closes #178.

Problem

An entry whose TTL had passed stayed in MemoryBackend.cache until the next sweep or get(). During that time, it was still visible:

• get_all_keys() and get_cache_data() listed it.
• clear_pattern(), clear_path() and delete_many() counted it as removed.

Redis never returns an expired key, so the same live data produced different numbers per backend. For example, CacheManager.clear_prefix() with one live key and one expired key returned 2 on memory and 1 on Redis.

Change

• _evict(), which clear_pattern, clear_path and delete_many share, still removes every matching entry but counts only the unexpired ones.
• get_all_keys() and get_cache_data() skip expired entries.

Effect on the monitoring routes

/cached-hits and /cached-records still compute is_expired and the expired_* counts. With the built-in backends, those now stay at zero apart from an entry that expires between the listing and the route reading the clock, as they already did on Redis. The route logic is kept for that case and for third-party backends that do list expired entries.

TestExpiredEntryMonitoring injected an expired entry into the memory backend's dict to exercise that logic. It now stubs get_cache_data() to return one, so it tests the route without relying on the old backend behaviour.

Tests

• test_memory_backend_enumeration_skips_expired_entries.
• test_memory_backend_does_not_count_expired_entries_as_cleared, parametrized over clear_pattern, clear_path and delete_many.
• test_clear_prefix_does_not_count_expired_keys, run on both memory and Redis to pin the parity.
• The new tests expire entries by rewriting expiry on memory and with PEXPIRE 1 on Redis, not by sleeping for a TTL.
• Mutation check: with the backend change reverted, exactly the five new memory cases fail. The Redis case passes either way, as expected.
• Full suite against live Redis and Memcached (CACHEX_REQUIRE_LIVE_SERVERS=1): 857 passed, 100% coverage.