fix(cache): serve uncached instead of 500 when the backend fails

[allen0099]

Closes #228.

Problem

@cache called backend.get() before the handler and backend.set() after it with no error handling. When the backend raised, every cached route returned 500, even when the handler had already produced a good response. A healthy Memcached does the same for any response over its 1 MB item size (MemcacheServerError: object too large for cache).

Fix

@cache fails open by default:

• A read error is logged as a warning on fastapi_cachex.cache and treated as a cache miss, so the handler runs.
• A write error is logged and the response is served unstored.
• @cache(fail_open=False) lets the error propagate as before, for routes that should fail loudly.

Only the backend calls themselves are wrapped. Handler exceptions, request-validation errors and decoration-time CacheXErrors are unaffected. A write failure does not stop reads: an entry stored earlier is still served.

Not in this PR

The issue lists two more ideas, now tracked separately:

• A BackendUnavailableError(CacheXError) that Redis and Memcached raise for transport errors (Raise a backend-independent BackendUnavailableError for transport failures #260). It would change the exception types that CacheManager, StateManager and sessions surface, so it needs its own design.
• Treating "object too large" as a skipped store for every caller (Handle Memcached "object too large" consistently outside @cache #261). For @cache it is now covered by fail-open. For CacheManager and set() callers, silently dropping a write is a behaviour choice of its own.

invalidate(), CacheManager, StateManager, CacheLock and sessions still raise backend errors. The docs say so.

Tests

tests/test_cache_backend_failure.py:

• test_failing_backend_serves_the_handler_response[get|set|both]: 200 with ETag and Cache-Control, the handler runs again, and the warning is logged.
• test_a_write_failure_does_not_hide_a_working_read
• test_fail_open_false_propagates_the_error[get|set]
• test_response_over_the_memcached_item_size_is_served_unstored: a live Memcached test with a 2 MB body.

Mutation check: with the old cache.py, the three fail-open cases and the live Memcached test fail. The opt-out and read-still-works tests pass on both, since they guard the new behaviour. The full suite passes against live Redis and Memcached: 939 passed, 99.96% coverage. Both docs builds pass with --strict, and the zh-TW anchor when-the-backend-fails matches the English one.

Docs

• HTTP_CACHING.md (EN and zh-TW): new "When the backend fails" section.
• BACKENDS.md (EN and zh-TW): the Memcached item size limit is added to the limitations.
• CHANGELOG: entry under ### Changed, since the default behaviour changes.
• CLAUDE.md: one line.