feat(redis)!: remove the clear_pattern prefix fallback and the encoding option

[allen0099]

Fixes #125. Fixes #126.

#125: clear_pattern prefix fallback
0.3.8 made clear_pattern() match the logical key. The retry with the prefix-stripped pattern, and its DeprecationWarning, are gone. A pattern that starts with key_prefix now matches only logical keys that themselves start with the prefix, the same as on every other backend. The docs in docs/BACKENDS.md (EN and zh-TW) are updated.

#126: encoding option, raw bytes

• The client is created with decode_responses=False. Replies go straight to decode_entry, which already accepted bytes, saving a decode/encode per hit. The Lua compare scripts behind delete_if_equals/expire_if_equals now receive exactly the bytes set() wrote.
• SCAN keys are decoded as UTF-8. A non-UTF-8 key under the prefix is skipped by get_all_keys/clear_path but still removed by clear().
• encoding= and decode_responses= passed to AsyncRedisCacheBackend raise TypeError. Otherwise they would fall through **kwargs into redis-py and switch decoding back on.
• decode_responses never warned in 0.3.9. It only accepted True, its default, and the migration guide now says so.
• RedisConfig drops the field and ignores an old encoding value like any other unknown field.
• Data 0.3.x stored with the default UTF-8 encoding reads back unchanged, and a test covers it.
• A caller-supplied pool with decode_responses=True still works.

Tests

• New: raw-bytes client, the TypeError, non-ASCII and non-UTF-8 content round-trips, old UTF-8 JSON entries, non-UTF-8 keys, and a decoding caller pool.
• Checks: pre-commit, mypy strict, the full suite including the Redis and Memcached suites (1517 passed, coverage 99%), and both strict docs builds.
• An independent review found no blocking issues; its wording and test nits are fixed in a7da2c8.

Changelog: changelog.d/125.removed.md, changelog.d/126.removed.md.