Valkey 9.0 supports numbered databases in cluster mode (cluster-databases
config).  Database selection is per-connection state, so a cluster client
must (re)issue SELECT on every node connection.  phpredis already does
exactly that for any socket with a nonzero dbNumber via the connect state
machine in redis_sock_server_open() and the reconnect path in
redis_check_eof() -- cluster sockets simply never had dbNumber set.

This adds:

* An eighth RedisCluster constructor argument (int $database = 0), plus a
  redis.clusters.database INI key for named clusters.
* RedisCluster::select() and RedisCluster::getdbnum().  select() eagerly
  validates the database against a single master (so servers without
  cluster database support fail cleanly with the error available from
  getlasterror()), then propagates lazily:  other sockets are disconnected
  and reselect the new database the next time they are used.
* Propagation of the configured database onto every node socket at
  creation (cluster_node_create, cluster_init_seeds, cluster_init_cache),
  which covers seeds, MOVED/ASK-discovered nodes, replicas and remaps.
* Database-segregated persistent connections:  the pool key and the
  non-pooled persistent id gain a :db<N> suffix when the database is
  nonzero, so a pooled stream left on db N can never be handed to a
  client expecting another database.  Keys for db 0 are unchanged.  This
  also fixes the same latent leak for standalone Redis when using the
  'database' option with persistent connections.
* Clearer errors when a backend rejects SELECT:  construct and
  first-command failures now surface the server error (e.g. "SELECT is
  not allowed in cluster mode") instead of generic communication errors.

When the database is 0 (the default) no SELECT is ever sent and pool keys
are byte-identical, so behavior with Redis, KeyDB and Valkey < 9 clusters
is completely unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Accept a database=N query parameter in session.save_path for the
rediscluster session handler, mirroring the standalone handler.  Requires
a server with database support in cluster mode (Valkey >= 9.0 with
cluster-databases > 1).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Covers both directions:  gated on a Valkey 9 cluster with databases
enabled, verify select(), the constructor argument, persistent pool
isolation, reconnect and the session handler save_path.  On servers
without support (Redis, KeyDB, Valkey < 9) verify that selecting a
nonzero database fails cleanly, leaves the connection usable and that
constructing with one raises a clear exception.

The standalone testPersistentDatabasePoolIsolation test fails without
the pool key fix, where a stream left on database 2 is handed back to a
client that asked for database 0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Add a make-cluster.sh -d option to spawn a cluster with numbered
databases, and pin one CI leg to Valkey 9.1.1 started with
cluster-databases 16 so the feature is actually exercised.  The
remaining valkey legs stay on 8.1.3 to keep covering servers without
database support in cluster mode.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

detectValkey() compared the 'executable' info field against the literal
string 'valkey', but that field holds a full path, so is_valkey was
never true and every minValkeyVersionCheck() gated test silently
skipped.  Match on server_name/valkey_version instead, which turns the
existing Valkey 9 geo polygon tests back on (both pass), along with the
new cluster database tests.

Also handle both reply shapes when reading cluster-databases from
CONFIG GET.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Without this a stale error from an earlier command could make a write
failure look like the server rejecting the database.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>