feat(redis)!: native async cluster Pub/Sub on redis-py 8

[Lancetnik]

Closes #3009. Targets 1.0.0 rather than main, because it raises the redis floor.

What this removes

Redis Cluster Pub/Sub has been running through a synchronous client since #2854, because redis.asyncio.cluster had no publish / pubsub at the time. redis-py 8.0.0 added them, so the workaround goes:

• _SyncPubSubProxy, sync_publish(), _get_sync_cluster() and the ThreadPoolExecutor in configs/state.py — along with the _sync_cluster / _thread_pool fields that the base ConnectionState carried for the cluster's sake
• RedisClusterFastProducer in publisher/producer.py, whose channel publish hopped to a thread on every message
• _make_channel_subscriber() and the subscriber() override in broker/cluster_broker.py, which existed only to monkeypatch start() on channel subscribers

Cluster now uses the same RedisFastProducer and the stock ChannelSubscriber.start() as standalone. BaseRedisFastProducer collapsed into RedisFastProducer since there is no longer a second producer to share with.

The credential_provider gap @fbartonoctopus reported on #3009 closes as a side effect: the hardcoded option whitelist in _get_sync_cluster() is gone, so every connection option reaches Channels, Lists and Streams through the one async client.

The dependency decision

The issue left this open. This PR takes option 2 — raise the floor to redis>=8.0.0 — rather than the runtime guard #2893 used. One supported path, and it lets faststream/redis/_compat.py go entirely: REDIS_V710 and REDIS_V720 are now always true, so the claim_min_idle_time version guard, the lib_name/lib_version fallback in _get_driver_info(), and the require_redis_v710 test marker all disappear with it.

This is breaking for every Redis user, not just cluster users — hence 1.0.0.

One real bug the live cluster surfaced

ClusterPubSub.execute_command reads nodes_manager.slots_cache directly instead of going through RedisCluster.execute_command, so it never triggers the client's lazy topology discovery. A channel subscriber is usually the first cluster operation an app performs, and it died on KeyError: <slot>. RedisClusterConnectionState.connect() now awaits client.initialize().

That makes connect() reach the network, which is what "connect" should mean — and it caught four tests that were only passing because the old connect was a no-op:

• test_start_stop_lifecycle ran RedisClusterBroker() against the default localhost:6379 with no cluster there; it is now connected-marked and points at the real cluster
• test_pipeline_warns published without starting the broker, which worked only because the sync client built itself lazily from _options; it now starts the broker, matching what standalone RedisBroker already required
• test_broker_gets_patched_attrs_within_cm and test_broker_with_real_doesnt_get_patched come from the shared BrokerTestclientTestcase and start the broker for real. The cluster overrides carry connected now, which moves them from test-redis-cluster-smoke, where no cluster runs, to test-redis-cluster-real

Checked during implementation

The issue flagged three things worth confirming, since #2893 predates current main:

• legacy_responses=True — not needed. redis-py 8 already defaults RedisCluster(legacy_responses=True), so RESP3 still yields RESP2-shaped responses and stream/list parsing is untouched. Passing it explicitly would be noise.
• refactor(redis): replace deprecated args: lib_name, lib_version to driver_in… #2956 (driver_info) — already landed; _get_driver_info() just loses its pre-7.2 branch.
• Sharded Pub/Sub — out of scope, as the issue says.

Validation

• ruff format / ruff check / codespell / mypy (1746 files) / pyright clean
• tests/brokers/redis + tests/docs/redis, in-memory: pass
• tests/brokers/redis -m redis against a live standalone Redis: pass
• tests/brokers/redis/cluster -m redis_cluster against a live 3-node cluster (redis 8, ports 7001-7003): pass