feat(backends): add set_if_absent and delete_if_equals for locks and slots

[allen0099]

Closes #62

What

Two atomic primitives on BaseCacheBackend, overridden by every built-in backend:

	set_if_absent(key, value, ttl=None) -> bool	delete_if_equals(key, expected) -> bool
Memory	check + store under the lock (expired = absent)	compare + delete under the lock
Redis	SET NX EX	decode + compare in Python, then a Lua script that deletes only if the key still holds the bytes compared
Memcached	ADD	GETS + CAS write with exptime -1 (immediately expired)
Base fallback	get-then-set (non-atomic, documented)	get-compare-delete (non-atomic, documented)

owner = CacheEntry(fingerprint="lock", content=secrets.token_bytes(16))
if await backend.set_if_absent(f"stream:{user_id}", owner, ttl=300):
    try:
        ...
    finally:
        await backend.delete_if_equals(f"stream:{user_id}", owner)

Why delete_if_equals too

#62 proposes releasing with delete. That frees the wrong holder when the first holder's entry has expired and someone else has claimed the key since — likely with long-lived SSE streams. Releasing only while the key still holds your token closes that.

Notes

• Memcached's classic DELETE takes no CAS token, so the release is a CAS write with a negative exptime. Verified against a real server: a stale CAS token fails, and a subsequent ADD claims the key.
• Comparison is on the decoded CacheEntry, so counters (stored as bare integers on Redis/Memcached) compare correctly.
• BackendProxy needs no change: callers already reach the backend via BackendProxy.get().
• Separate commit: test_the_repository_changelog_can_be_released hardcoded v0.3.4 and ### Security, so it has been failing on master since 0.3.5 was released. It now reads the latest release from the file.

Testing

• Per backend: first-writer-wins, 20-way concurrency has exactly one winner, mismatch does not delete, counter comparison, and the Add an atomic set-if-absent (SET NX EX) to BaseCacheBackend for lock / slot acquisition #62 scenario (release after expiry keeps the new holder).
• Redis/Memcached race tests overwrite the key between compare and delete. Swapping either implementation for a plain delete fails exactly those two tests.
• Locally against live Redis/Memcached: 665 passed, 1 skipped (live-server gate), coverage 100%; ruff, mypy --strict and pre-commit clean.